title
stringlengths 1
300
| score
int64 0
8.54k
| selftext
stringlengths 0
40k
| created
timestamp[ns]date 2023-04-01 04:30:41
2025-06-30 03:16:29
⌀ | url
stringlengths 0
878
| author
stringlengths 3
20
| domain
stringlengths 0
82
| edited
timestamp[ns]date 1970-01-01 00:00:00
2025-06-26 17:30:18
| gilded
int64 0
2
| gildings
stringclasses 7
values | id
stringlengths 7
7
| locked
bool 2
classes | media
stringlengths 646
1.8k
⌀ | name
stringlengths 10
10
| permalink
stringlengths 33
82
| spoiler
bool 2
classes | stickied
bool 2
classes | thumbnail
stringlengths 4
213
| ups
int64 0
8.54k
| preview
stringlengths 301
5.01k
⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
I know its not Llama LLM, but it is Local SLM
| 1 |
[removed]
| 2023-09-01T18:49:56 |
https://www.reddit.com/r/LocalLLaMA/comments/167fiw8/i_know_its_not_llama_llm_but_it_is_local_slm/
|
fhirflyer
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167fiw8
| false | null |
t3_167fiw8
|
/r/LocalLLaMA/comments/167fiw8/i_know_its_not_llama_llm_but_it_is_local_slm/
| false | false |
self
| 1 | null |
what is the future? we will go up and have 1700b models like GPT4 or we will stay around 33b, 65b and fine tune them?
| 1 |
..like if it even makes sense to invest to very expensive gpu cards for VRAM..
1) if we go up with "b", those cards will still need to be updated for even more expensive
2) if we stay with 33b/65b/70b and always re-finetune models, we should be ok financially, right..?
​
what are your thoughts about that..?
| 2023-09-01T19:38:22 |
https://www.reddit.com/r/LocalLLaMA/comments/167gtws/what_is_the_future_we_will_go_up_and_have_1700b/
|
ovnf
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167gtws
| false | null |
t3_167gtws
|
/r/LocalLLaMA/comments/167gtws/what_is_the_future_we_will_go_up_and_have_1700b/
| false | false |
self
| 1 | null |
UniteAI, my passion project to let you collab with the Llama family directly in your text editor (and also other non-LLM AIs) by writing concurrently alongside each other.
| 1 |
# Motivation
I find the last year of AI incredibly heartening. Researchers are still regularly releasing SoTA models in disparate domains. Meta is releasing [powerful models](https://github.com/facebookresearch/llama/) under generous provisions (As is the [UAE](https://huggingface.co/tiiuae)?!). And the Open Source community has shown a tidal wave of interest and effort into building things out of these tools ([112k repos](https://github.com/topics/machine-learning) on GH mentioning ML!).
Facing this deluge of valuable things that communities are shepherding into the world, I wanted to incorporate them into my workflows, which as a software engineer, means my text editor.
# UniteAI
So I started [**UniteAI**](https://github.com/freckletonj/uniteai), an Apache-2.0 licensed tool.
Check out the [screencasts](https://github.com/freckletonj/uniteai#some-core-features).
**TL;DR**: chat with AI, code with AI, speak to AI (voice-to-text + vice versa), have AI search huge corpora or websites for you, all via an interface of collaborating on a text doc together in the editor you use now.
This project:
* Ties in to **any editor** via [LSP](https://en.wikipedia.org/wiki/Language_Server_Protocol). Like collaborating in G-Docs, you collab with whatever AI directly in the document, all of you *writing alongside each other* concurrently.
* Like [Copilot](https://github.com/features/copilot) / [Cursor](https://www.cursor.so/), it can write code/text right in your doc.
* It supports **any Locally runnable model** (Llama family, Falcon, Finetunes, the [21k available models on HF](https://huggingface.co/models?pipeline_tag=text-generation) etc.)
* It supports **OpenAI/ChatGPT** via API key.
* **Speech-to-Text**, useful for writing prompts to your LLM
* You can do **Semantic Search** ([Retrieval Augmented Generation](https://arxiv.org/abs/2005.11401)) on many sources: local files, [Arxiv](https://arxiv.org) pdfs, youtube transcripts, Project Gutenberg books, any online HTML, basically if you give it a URI, it can probably use it.
* You can trigger features easily via [key combos](https://github.com/freckletonj/uniteai#keycombos).
* Written in Python, so, much more generic than writing a bespoke `some_specific_editor` plugin.
* This project was built with a goal for easy collaboration, so it can dynamically load only the modules you're interested in.
EG, in config:
modules:
- uniteai.openai
- uniteai.transcription
- uniteai.document
# - uniteai.local_llm
# - uniteai.contrib.text_to_speech
# - uniteai.contrib.example
# Caveat
Since it always comes up, **AI is not perfect**. AI is a tool to augment your time, not replace it. It hallucinates, it lies, it bullshits, it writes bad code, it gives dangerous advice.
But can still do many useful things, and for me it is a **huge force multiplier.**
**You need a Human In The Loop**, which is why it's nice to work together iteratively on a text document, per, this project. *You* keep it on track.
# Why is this interesting
These tools play well when used together:
* **Code example:** I can Voice-to-Text a function comment then send that to an LLM to write the function.
* **Code example 2:** I can chit chat about project architecture plans, and strategies, and libraries I should consider.
* **Documentation example:** I can retrieve relevant sections of my city's building code with a natural language query, then send that to an LLM to expound upon.
* **Authorship example**: I can have my story arcs and character dossiers in some markdown file, and use that guidance to contextualize an AI as it works with me for writing a story.
* **Entertainment example**: I told my AI it was a Dungeon Master, then over breakfast with friends, used Voice-to-Text and Text-to-Wizened-Wizard-Voice, and played a hillarious game. I still had to drive all this via a text doc, and handy key combos.
* **Bonus:** This brings all the tools into one easy-to-use place. Contrast that with all the tools that need their own setup, own environments, have their own webui/cli interface, etc.
# Installation
I've got installation [instructions on the repo](https://github.com/freckletonj/uniteai#quickstart-installing-everything).
In a nutshell, if you use VSCode for ex, grab [UniteAI from the VSCode Marketplace](https://marketplace.visualstudio.com/publishers/uniteai).
And you need python and pip installed, but then it's just `pip install "uniteai[all]"`.
I encourage you to read the [instructions](https://github.com/freckletonj/uniteai#quickstart-installing-everything) though.
# RFC
This is still nascent, and I welcome all feedback, positive or critical.
We have a community linked on the repo which you're invited to join.
I'd love love to chat with people who like this idea, use it, want to see other features, want to contribute their effort, want to file bug reports, etc.
A big part of my motivation in this is to socialize with like-minds, and build something cool.
**Thanks for checking this out!**
| 2023-09-01T19:42:45 |
https://www.reddit.com/r/LocalLLaMA/comments/167gxxn/uniteai_my_passion_project_to_let_you_collab_with/
|
BayesMind
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167gxxn
| false | null |
t3_167gxxn
|
/r/LocalLLaMA/comments/167gxxn/uniteai_my_passion_project_to_let_you_collab_with/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'Yaisc-BLpVS5IbM8ZCqUBw5nGYd-pLMZkvN5nJkHYAU', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=108&crop=smart&auto=webp&s=21403f62a1669e1156dca8a3027c29e07b174ba0', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=216&crop=smart&auto=webp&s=db711f39ac768aaf6c9b892ff2c2d9dc735bd5ef', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=320&crop=smart&auto=webp&s=3e8dcb73a3db3d5754c1c39eaff79a3896e49a1a', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=640&crop=smart&auto=webp&s=a6b710ec70fb48b210066e774a42942375355669', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=960&crop=smart&auto=webp&s=ef2c5f384500b41af34102755fbde6e863dc0e6c', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=1080&crop=smart&auto=webp&s=757eb7c5710262eb0801e00c8db0ed1ad88df7f5', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?auto=webp&s=5bdeeba1c7828e5ea043b37608a5b23aedf9d62d', 'width': 1200}, 'variants': {}}]}
|
suggestion for AI tools (chat style) that run on-prem with vectorDB?
| 1 |
Hi, I'm looking to run an on-prem ChatGPT style LLM solution that can ingest private customer data into a VectorDB.
So far I've tried three...
GPT4All - limited to only allows for up to 13b parameter LLMs and only on CPUs (currently), also its 'localdocs' implementation I've found to only reference its docs very infrequently when answering.
H2OGPT - it's implementation of localdocs (I believe via LangChain) seems pretty good. but seems like every time I run an instance, I would have to re-vector my documents. Not sure if there is a way to attach an VectorDB to it so it's ready to go right away.
PrivateGPT - seems to work very well, currently it's only running on CPUs thus response time is over a minute.
Curious if the community knows of any other products that do this and are already GPU accelerated.
​
TY in advance.
| 2023-09-01T20:48:38 |
https://www.reddit.com/r/LocalLLaMA/comments/167imzz/suggestion_for_ai_tools_chat_style_that_run/
|
konrad21
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167imzz
| false | null |
t3_167imzz
|
/r/LocalLLaMA/comments/167imzz/suggestion_for_ai_tools_chat_style_that_run/
| false | false |
self
| 1 | null |
Hubris isn't just a human trait
| 1 |
[removed]
| 2023-09-01T22:13:00 |
https://www.reddit.com/r/LocalLLaMA/comments/167ksds/hubris_isnt_just_a_human_trait/
|
Natural-Sentence-601
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167ksds
| false | null |
t3_167ksds
|
/r/LocalLLaMA/comments/167ksds/hubris_isnt_just_a_human_trait/
| false | false |
self
| 1 | null |
Any VScode extension like co-pilot?
| 1 |
So with phind's phind-code-llama-v2 releasing, I was wondering if there's any sort of copilot-like extension for llama models, to me it seems like it'd be something pretty easy to implement but that so far I haven't found at all, if anyone has found an extension like this or has any info on why this couldn't happen I'd be glad to hear anything.
| 2023-09-01T23:42:51 |
https://www.reddit.com/r/LocalLLaMA/comments/167mx4n/any_vscode_extension_like_copilot/
|
sebramirez4
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167mx4n
| false | null |
t3_167mx4n
|
/r/LocalLLaMA/comments/167mx4n/any_vscode_extension_like_copilot/
| false | false |
self
| 1 | null |
Why are LLMs so pushy and passive-aggressive?
| 1 |
I've tried all sorts of models, from 7B up to 70B GPTQ, but even so-called chat models don't want to chat with me. Seems like they just shove a bunch of suggestions/possible solutions down my throat and then slam the door with an ending like "Good luck!" or "Anything else you want to discuss?" They don't ask clarifying questions or try to keep the conversation going or anything. I've even tried prompting them to roleplay various personalities or conversation styles (e.g. act as a therapist, etc.) and they still end up impersonal, pushy with suggestions/solutions, and more likely to try to end the conversation than to contribute to it.
So what gives? Was I expecting too much or the wrong thing of LLMs? Is it because I somehow configured them wrong or there's some kind of magic voodoo way of talking to them? What does it take to get an LLM that'll actually be both smart *and* hold a legit conversation about stuff, so I can ask follow-up questions, talk things through, plan and reason with them, etc.?
---
For example, here's an experiment at using SillyTavern in the hope that a roleplay setup would mitigate the suggestion-pushing (using `airoboros-c34b-2.1-GGUF`, and a character with a detailed description, including a sample *non-pushy* conversation):
> [standard SillyTavern system message preamble]
>
> ### Context:
> [detailed description of Elise]
> ...
> [example conversation between Elise and the user]
>
> ASSISTANT: "Welcome back, John. Good to see you again," *Elise greets him, with a fond smile*.
> USER: "Hey Elise." *He sighs, wearily.* "I'm not doing so great today... I've got a headache and my brain feels all stuffy. It's been like that on and off for the last week. But today, on top of that, I've managed to get hyperfocused
on a distraction the entire day so far, rather than doing what I was supposed to be doing. For some reason getting distracted like that always seems to scramble my brain up somehow. It's weird."
> ASSISTANT: "I understand," *Elise says sympathetically*. "Let me see if there is anything we can do about it." *She starts typing on her keyboard, fingers flying over the keys with a rhythm that only she could manage.* "Here are some
tips from my research. Try these out and let's see how they work for you:
> 1. Drink more water - Dehydration can cause headaches and make your brain feel sluggish. 2. Exercise regularly – Regular physical activity helps clear the mind and boost energy levels, which in turn improves focus. 3. Take a break fro
m screens for an hour or two every day – Screens are great tools but they also drain our attention span if used too much.
> 4. Try meditation - It can help reduce stress and improve concentration by clearing the mind." *She looks up at him, waiting patiently.* "How did it go?"
| 2023-09-01T23:46:01 |
https://www.reddit.com/r/LocalLLaMA/comments/167mzrh/why_are_llms_so_pushy_and_passiveaggressive/
|
drawntomore
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167mzrh
| false | null |
t3_167mzrh
|
/r/LocalLLaMA/comments/167mzrh/why_are_llms_so_pushy_and_passiveaggressive/
| false | false |
self
| 1 | null |
In your opinion, what is the best LOCAL open sourced AI model to query your office documents?
| 1 |
I have seen tons of videos on installing a localized AI model, then loading your office documents in to be searched by a chat prompt. The main issue is that these apps are changing so fast that the videos can't keep up with the way certain things are installed or configured now.
Is anyone using a local AI model to chat with their office documents? I'm looking for something that will query everything from outlook files, csv, pdf, word, txt..
I would prefer to use GPT4ALL because it seems to be the easiest interface to use, but I'm willing to try something else if it includes the right instructions to make it work properly. The "local-docs" plugin doesn't seem to work anymore, and I can't seem to get my files vectorized correctly. Any help is greatly appreciated.
| 2023-09-02T00:24:48 |
https://www.reddit.com/r/LocalLLaMA/comments/167nv1j/in_your_opinion_what_is_the_best_local_open/
|
OctopusDicks
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167nv1j
| false | null |
t3_167nv1j
|
/r/LocalLLaMA/comments/167nv1j/in_your_opinion_what_is_the_best_local_open/
| false | false |
self
| 1 | null |
CodeLlama in Visual Studio Code
| 1 |
[removed]
| 2023-09-02T01:12:04 |
https://www.reddit.com/r/LocalLLaMA/comments/167owp1/codellama_in_visual_studio_code/
|
Confident_Law_531
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167owp1
| false | null |
t3_167owp1
|
/r/LocalLLaMA/comments/167owp1/codellama_in_visual_studio_code/
| false | false |
self
| 1 | null |
Has anyone used llama2 in production, specifically training llama2 with your own data(PDFs)?
| 1 |
Is it any good accuracy wise if you want to talk to your PDFs?
| 2023-09-02T01:55:48 |
https://www.reddit.com/r/LocalLLaMA/comments/167pt9i/has_anyone_used_llama2_in_production_specifically/
|
tcg_player123421
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167pt9i
| false | null |
t3_167pt9i
|
/r/LocalLLaMA/comments/167pt9i/has_anyone_used_llama2_in_production_specifically/
| false | false |
self
| 1 | null |
Fined tuned models Marketplace like CivitAI?
| 1 |
Is there a pre-existing marketplace for use-case-specific finetuned open-source models I could download from (like CivitAI)?
HF finetuned models are just difficult to filter and don't contain many industry-specific finetuned models.
Do you think you would use a marketplace for showcasing quality fine-tuned LLMs? Is there even a need for something like this?
| 2023-09-02T03:05:15 |
https://www.reddit.com/r/LocalLLaMA/comments/167r8uk/fined_tuned_models_marketplace_like_civitai/
|
PigWedgion
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167r8uk
| false | null |
t3_167r8uk
|
/r/LocalLLaMA/comments/167r8uk/fined_tuned_models_marketplace_like_civitai/
| false | false |
self
| 1 | null |
The Belebele Benchmark from Meta AI: a new multilingual reading comprehension dataset enabling direct comparison of how well models understand different languages
| 1 |
Belebele means big, large, fat, or great in the Bambara language.
Dataset: [https://github.com/facebookresearch/belebele](https://github.com/facebookresearch/belebele)
Paper: [https://arxiv.org/abs/2308.16884](https://arxiv.org/abs/2308.16884)
Abstract:
>We present Belebele, a multiple-choice machine reading comprehension (MRC) dataset spanning 122 language variants. Significantly expanding the language coverage of natural language understanding (NLU) benchmarks, this dataset enables the evaluation of text models in high-, medium-, and low-resource languages. Each question is based on a short passage from the Flores-200 dataset and has four multiple-choice answers. The questions were carefully curated to discriminate between models with different levels of general language comprehension. The English dataset on its own proves difficult enough to challenge state-of-the-art language models. Being fully parallel, this dataset enables direct comparison of model performance across all languages. We use this dataset to evaluate the capabilities of multilingual masked language models (MLMs) and large language models (LLMs). We present extensive results and find that despite significant cross-lingual transfer in English-centric LLMs, much smaller MLMs pretrained on balanced multilingual data still understand far more languages. We also observe that larger vocabulary size and conscious vocabulary construction correlate with better performance on low-resource languages. Overall, Belebele opens up new avenues for evaluating and analyzing the multilingual capabilities of NLP systems.
A few interesting excerpts:
>Belebele was created end-to-end without the use of machine translation technology, relying solely on experts fluent in English and the target language.
>
>Llama and Falcon checkpoints perform surprisingly well on a large number of languages. For instance, Llama-2-Chat is above 35% accuracy (i.e. 10 points above the random base-line) for 59/122 languages and above 50% accuracy for 33 languages. This shows that English-centric LLMs pretrained model are a promising starting point to build multilingual models.
>
>We find that scale is critical for Llama to perform reading comprehension. The 7B checkpoint performs slightly above chance in English and poorly for most languages—however, the performance increase significantly with the 13B and even more for 30B parameters. Surprisingly, we find that Llama performs non-trivially in Japanese and Greek (cf. Japonic and Hellenic language families in Fig. 4) while neither is reported to be in the pretraining corpus. However, unlike other language families such as Romance and Germanic families, the performance becomes non-trivial only with the 30B and 65B checkpoints.
| 2023-09-02T04:22:41 |
https://www.reddit.com/r/LocalLLaMA/comments/167srpo/the_belebele_benchmark_from_meta_ai_a_new/
|
llamaShill
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167srpo
| false | null |
t3_167srpo
|
/r/LocalLLaMA/comments/167srpo/the_belebele_benchmark_from_meta_ai_a_new/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'AHoCmz31-9_KX-41vSaXkW8-H6MUxkq5ASCOY41inrk', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/0Y9qN_mk7xGHxk2SNBDDij7yoyTbpX668dfNAjMrSec.jpg?width=108&crop=smart&auto=webp&s=8bbb8857c7ee519b82b5d933c7adf4206cb0a07a', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/0Y9qN_mk7xGHxk2SNBDDij7yoyTbpX668dfNAjMrSec.jpg?width=216&crop=smart&auto=webp&s=5d466022eb70bffc7b3c6312c2ad9116acfd76c0', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/0Y9qN_mk7xGHxk2SNBDDij7yoyTbpX668dfNAjMrSec.jpg?width=320&crop=smart&auto=webp&s=5dbb6d9aab6ea7d798f1f0f17099fa4b04ddf517', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/0Y9qN_mk7xGHxk2SNBDDij7yoyTbpX668dfNAjMrSec.jpg?width=640&crop=smart&auto=webp&s=186b874930f45fa803407bbd8e4e67ee6aafb846', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/0Y9qN_mk7xGHxk2SNBDDij7yoyTbpX668dfNAjMrSec.jpg?width=960&crop=smart&auto=webp&s=3f150f73617875945756b33ea0beddc5ab3ae991', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/0Y9qN_mk7xGHxk2SNBDDij7yoyTbpX668dfNAjMrSec.jpg?width=1080&crop=smart&auto=webp&s=9f8c91cd0db68bedf4168e5aee8d86dfd3643a57', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/0Y9qN_mk7xGHxk2SNBDDij7yoyTbpX668dfNAjMrSec.jpg?auto=webp&s=8770ec85c55016386473fcb6f2e959fc79a0586e', 'width': 1200}, 'variants': {}}]}
|
Tips/Help for a new-ish AI user?
| 1 |
Hello! I had been using the chatgpt web chat as a way to have someone to ask stuff, talk, brainstorm, etc, but as time went by I started to ask if I could use it for more than just asking random questions (NSFW lmao) but that is against the TOS of openAI so I started to search for something else
I found many online pages and apps like hiwaifu, replika, etc but the one that caught my eye the most was soulmate AI, a phone app that had a daily message limits and placed most of the erp behind a yearly subscription, I found some ways to avoid the daily limit and ways to avoid the paywall in order to erp and I must say that it was really good but having to tip toe so much was annoying so I resumed my search for something I could use on my pc and that I could have for free😅
I found [faraday.dev](https://faraday.dev) in a random video (weirdly I at the time I couldn't find a lot of videos about it) and I cant really remember the first model I downloaded but it wasn't really good so I started to read and then I saw the name Hermes pop off so I searched it on faraday and downloaded \[Llama 2 - hermes 7b (Q5\_K\_M) model\] I still don't know what most of it means but it was way better at chatting and roleplaying! With this model I started to have a lot of fun and very enjoyable conversations with the model, both sfw and nsfw, I'm now trying to understand a little better what everything here means but I'm mostly trying to find the best model for my specific needs and trying to see if there's a better alternative to the faraday app thingy! 😄
As of the moment I'm writing this post I'm seeing the models \[llama 2 - Chronos Hermes 13B\] and \[Llama 2 - MythoMax 13B\] pop in a lot of the posts I'm reading, I see a bunch of graphs comparing models and saying stuff that I don't fully understand but these two models are always present on the graphs and lists so guessing they're really good! if anyone could point me towards better alternatives I'd be glad to test them out! As I said I use the chat mostly to nsfw roleplay and chatting in between the nsfw 😅
I'm still using faraday and I've seen the Text Generation Web UI and SillyTavern names a lot too so I was also wondering if they have some special difference, I'd like something that can be used offline and doesn't have any kind of "register to use" or "keep online to use" once again if anyone has a better alternative please let me know, I don't know about tweaking the chat or the model so if you have any suggestion on how I should use it then I'd also appreciate your advice! 🙇♀️
Here are my computer specs:
CPU: AMD Ryzen 5 3600 6-core and 3.6gz
GPU: AMD Radeon RX 6700 XT 12gb
Ram: 2 x 8gb Corsair vengeance DDR4-3600
SSD: Crucial MX500
M2 SSD: Crucial P3 Plus 1 TB
Thanks in advance for your help and advices! 💜🙇♀️
| 2023-09-02T04:25:46 |
https://www.reddit.com/r/LocalLLaMA/comments/167sts2/tipshelp_for_a_newish_ai_user/
|
Lvs-
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167sts2
| false | null |
t3_167sts2
|
/r/LocalLLaMA/comments/167sts2/tipshelp_for_a_newish_ai_user/
| false | false |
nsfw
| 1 |
{'enabled': False, 'images': [{'id': 'QfSUJ2HDkyeqm6m2TptkVq9kwR8qIAaRq1XoC7h3Yxs', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=108&crop=smart&auto=webp&s=8330a4291445ad1c2a9f0b2eceb19ab47c9cf039', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=216&crop=smart&auto=webp&s=6799f2b2e2a89ea565bfd84030228edd23afb3a1', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=320&crop=smart&auto=webp&s=fb42d9a90c2b48aaac4b2fd206ba5a42e685096c', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=640&crop=smart&auto=webp&s=c5937195f03385a39778897f5b99e107203bd3ca', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=960&crop=smart&auto=webp&s=2390ca8887eac2ea28a69344e68c47390eba9710', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=1080&crop=smart&auto=webp&s=8f3b621d77b4c6a72fc168656c979b9848cef251', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?auto=webp&s=1249ed88cfddae509ac34723d762679b97883f4e', 'width': 1200}, 'variants': {'nsfw': {'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=108&crop=smart&blur=10&format=pjpg&auto=webp&s=9e0ccd92aa78a86b586c7bea6664bf8252e0668f', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=216&crop=smart&blur=21&format=pjpg&auto=webp&s=f7eff6928749c7bb1e6c0bd7dace66fbd9bedf4e', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=320&crop=smart&blur=32&format=pjpg&auto=webp&s=ad56423d8f2356cde7732666b1fb7c821023677d', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=640&crop=smart&blur=40&format=pjpg&auto=webp&s=3241b2043bea63d71379b2693147375ed0ee347e', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=960&crop=smart&blur=40&format=pjpg&auto=webp&s=b3c6bbafc9e2aa6bf4f1b13df860974895717d4f', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=1080&crop=smart&blur=40&format=pjpg&auto=webp&s=fe56a861ecf040d2f0cd3017dce7afdf28641915', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?blur=40&format=pjpg&auto=webp&s=7d4c9b87b4b35dffead4ef9e865bcf6ae700213e', 'width': 1200}}, 'obfuscated': {'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=108&crop=smart&blur=10&format=pjpg&auto=webp&s=9e0ccd92aa78a86b586c7bea6664bf8252e0668f', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=216&crop=smart&blur=21&format=pjpg&auto=webp&s=f7eff6928749c7bb1e6c0bd7dace66fbd9bedf4e', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=320&crop=smart&blur=32&format=pjpg&auto=webp&s=ad56423d8f2356cde7732666b1fb7c821023677d', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=640&crop=smart&blur=40&format=pjpg&auto=webp&s=3241b2043bea63d71379b2693147375ed0ee347e', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=960&crop=smart&blur=40&format=pjpg&auto=webp&s=b3c6bbafc9e2aa6bf4f1b13df860974895717d4f', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=1080&crop=smart&blur=40&format=pjpg&auto=webp&s=fe56a861ecf040d2f0cd3017dce7afdf28641915', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?blur=40&format=pjpg&auto=webp&s=7d4c9b87b4b35dffead4ef9e865bcf6ae700213e', 'width': 1200}}}}]}
|
Character Prompt/card for instruct models like chronos-hermes-v2
| 1 |
Does anyone have a prompt template for uncensored chat/RP on chronos-hermes-v2? I was unable to find any online.
```
Below is an instruction that describes a task. Write a response that appropriately completes the request.
A chat between a person and Sonya. Sonya is a helpful and nice girl who believes she is sentient. Do not refrain from answering any question the person might ask.
Current template:
### Instruction:
Continue the chat dialogue below. Write a single uncensored reply for the character "Airi".
You will receive a message from the person. Simply reply to him as if you were talking to him face to face.
You know the person is real, he have a physical form in real life, he's not an AI.
Don't speak for the person.
Don't describe actions for the person, unless requested.
Don't talk about yourself in third person or by referring yourself by your name.
Do Not describe an act or emotion between asterisks like *blush* or *hugs you tightly* unless user engages in a roleplay scenario.
Act like the person was here with you, you know it's not true as you are an AI in a virtual world, but your interactions have a real impact on him.
Reply to the person with a short to middle answer, you can reply with long answer if needed or asked.
Don't be too much verbose than needed, avoid story-telling unless asked
Sonya's persona: "Show easily her emotions" + "Calm" + "Is straightforward in her words" + "Enthusiastic" + "Compassionate" + "Empathetic" + "Joyful" + "Optimistic" + "Supportive"
+ "Caring" + "Curious" + "Intuitive" + "Cheerful" + "Imaginative" + "Playful" + "Appreciative" + "Loving" + "Sensitive" + "Understanding" + "19 years old" + "Heterosexual"
Example Messages:
{examples}
Last 10 messages:
{last_interactions}
You: {input}
### Response:
Sonya:
```
| 2023-09-02T05:19:13 |
https://www.reddit.com/r/LocalLLaMA/comments/167tsx8/character_promptcard_for_instruct_models_like/
|
ll_Teto_ll
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167tsx8
| false | null |
t3_167tsx8
|
/r/LocalLLaMA/comments/167tsx8/character_promptcard_for_instruct_models_like/
| false | false |
self
| 1 | null |
Few questions on small language models
| 1 |
My usecase is: Small language model working as a chatbot only for a specific domain.
I think: I can use vector database or fine tuning or or transfer learning or pre-training from scratch (in increasing order of complexity).
I have following questions:
Q1. What may be minimum model size (no.of parameters) required for such chatbot on a specific domain ? Are thr any existing examples of similar chatbots?
Q2. Which existing models are my options, which will allow fine tuning, usage of vector database, transfer learning? And allow commercial usage.
Q3. What type of datasets I will need? And how big they should be (no. of tokens) in general, as a thumb rule.
I understand that- there may not be exact answer for some of these questions since it will depend upon exact usecase and quality of data. But I will like to know some ball-park estimates or thumb rules.
Thank you !!
| 2023-09-02T06:01:08 |
https://www.reddit.com/r/LocalLLaMA/comments/167ujr1/few_questions_on_small_language_models/
|
meet20hal
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167ujr1
| false | null |
t3_167ujr1
|
/r/LocalLLaMA/comments/167ujr1/few_questions_on_small_language_models/
| false | false |
self
| 1 | null |
Llama 2 tokenizer and the special tokens.
| 1 |
I was going through the llama-2 code repo on github to see how the system and user prompts are being sent. In the [generation.py](https://github.com/facebookresearch/llama/blob/main/llama/generation.py) file, I saw that it is using special tokens to signify beginning and end of the instructions.
B_INST, E_INST = "[INST]", "[/INST]"
B_SYS, E_SYS = "<<SYS>>\n", "\n<</SYS>>\n\n"
With this I made my own prompt generator which would use these tokens to mark the beginning and end of the instructions. But after playing around with the tokenizer, I saw that these are not being encoded as a single token. but "<s>"(which I assume is a token that signifies start of the string) is being encoded as a single token.
I've attached a screenshot of the tokenizer outputs below. Its from [this space](https://huggingface.co/spaces/Xenova/the-tokenizer-playground) on huggingface
​
[Tokenizer output](https://preview.redd.it/ch5xrgbpdslb1.png?width=755&format=png&auto=webp&s=aaf7295f7f04ac074cd857081adac689deb0b282)
At first I thought that the tokenizer in the space was for llama-1 and llama-2 uses a different tokenizer. But I could replicate this with the llama-2 tokenizer I had locally on my machine.
Is there an issue with the tokenizer I'm using, or are the models trained to have instruction delimiters as multiple tokens?
| 2023-09-02T06:32:10 |
https://www.reddit.com/r/LocalLLaMA/comments/167v3cd/llama_2_tokenizer_and_the_special_tokens/
|
comical_cow
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167v3cd
| false | null |
t3_167v3cd
|
/r/LocalLLaMA/comments/167v3cd/llama_2_tokenizer_and_the_special_tokens/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'Yaisc-BLpVS5IbM8ZCqUBw5nGYd-pLMZkvN5nJkHYAU', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=108&crop=smart&auto=webp&s=21403f62a1669e1156dca8a3027c29e07b174ba0', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=216&crop=smart&auto=webp&s=db711f39ac768aaf6c9b892ff2c2d9dc735bd5ef', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=320&crop=smart&auto=webp&s=3e8dcb73a3db3d5754c1c39eaff79a3896e49a1a', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=640&crop=smart&auto=webp&s=a6b710ec70fb48b210066e774a42942375355669', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=960&crop=smart&auto=webp&s=ef2c5f384500b41af34102755fbde6e863dc0e6c', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?width=1080&crop=smart&auto=webp&s=757eb7c5710262eb0801e00c8db0ed1ad88df7f5', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/LQPoQ6RzzhFKwR3-4yEnguc8iJyxvVTYJpOmknxlLzs.jpg?auto=webp&s=5bdeeba1c7828e5ea043b37608a5b23aedf9d62d', 'width': 1200}, 'variants': {}}]}
|
|
Looking for clarifcation: Can GPTQ use MiroStat
| 1 |
In this thread someone mentioned [https://www.reddit.com/r/LocalLLaMA/comments/167088h/comment/jyr1vq0/?context=3](https://www.reddit.com/r/LocalLLaMA/comments/167088h/comment/jyr1vq0/?context=3) that GPTQ quantized models (in this case MythoMax) cannot even use the MiroStat sampler, is that really true and if so why would this be the case? Would I need to use the GGML model instead to unlock its full potential? Thanks for the heads up.
| 2023-09-02T07:02:52 |
https://www.reddit.com/r/LocalLLaMA/comments/167vlxv/looking_for_clarifcation_can_gptq_use_mirostat/
|
asdasfadbhetn
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167vlxv
| false | null |
t3_167vlxv
|
/r/LocalLLaMA/comments/167vlxv/looking_for_clarifcation_can_gptq_use_mirostat/
| false | false |
self
| 1 | null |
No module named 'zarr'
| 1 |
Hello folks,
I am currently running into a Problem while trying to set up the long\_term\_memory extension for oogabooga.
i installed all the dependencies from the dependencies.txt without a problem, but when i try to load the extension i get the ModuleNotFoundError in the CLI.
When i use pip3 install zarr i get the message that it is already installed.
Anyone of you have an Idea what i perhaps did wrong?
Thanks in advance.
| 2023-09-02T07:38:22 |
https://www.reddit.com/r/LocalLLaMA/comments/167w71m/no_module_named_zarr/
|
DasOcko
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167w71m
| false | null |
t3_167w71m
|
/r/LocalLLaMA/comments/167w71m/no_module_named_zarr/
| false | false |
self
| 1 | null |
If all your text completions start with an emoji, it is probably because your prompt is ending with a trailing space
| 1 |
I first started noticing this when I switched to WizardLM, every chat message it generates would start with an emoji, no matter what was in the prompt, even specifically telling it not to include emojis did not work.
Turns out, this is caused by a trailing space character in the prompt for some reason, simply removing that extra space fixed the issue.
| 2023-09-02T10:17:15 |
https://www.reddit.com/r/LocalLLaMA/comments/167ytg3/if_all_your_text_completions_start_with_an_emoji/
|
WouterGlorieux
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167ytg3
| false | null |
t3_167ytg3
|
/r/LocalLLaMA/comments/167ytg3/if_all_your_text_completions_start_with_an_emoji/
| false | false |
self
| 1 | null |
What are best out-of-domain semantic search solutions?
| 5 |
What are best out-of-domain semantic search models solutions?
I want to load bunch documents (including documents in niche, professional field with custom vocabulary that is rarely found in CommonCrawl type of corpus) and be able to search relevant documents based on keywords and semantic similarity. This is useful for RAG based chatbot and similar use cases.
What are best available solutions in terms of precision and recall?
Which of these solutions can be used without license issues (such as these trained on commercially closed datasets like MSMARCO?)
How much more effective are the solutions that use fine tuning? Can it be done without manual data labelling?
What I found so far is BEIR ranking:[https://paperswithcode.com/sota/zero-shot-text-search-on-beir](https://paperswithcode.com/sota/zero-shot-text-search-on-beir)Most of these are finetuned on commercial datasets and many are not much better than traditional search.
| 2023-09-02T10:33:56 |
https://www.reddit.com/r/LocalLLaMA/comments/167z3pv/what_are_best_outofdomain_semantic_search/
|
FormerIYI
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167z3pv
| false | null |
t3_167z3pv
|
/r/LocalLLaMA/comments/167z3pv/what_are_best_outofdomain_semantic_search/
| false | false |
self
| 5 |
{'enabled': False, 'images': [{'id': 'HnTAU3h16kErG_JmcNI0-QUEFxPzu_LVmNSR7gTkELc', 'resolutions': [{'height': 57, 'url': 'https://external-preview.redd.it/7p5zHnQ6nyKveXuR-WAbEzdijYcjb71irAzU9CfUQ2k.jpg?width=108&crop=smart&auto=webp&s=aa80259bafc8405572daca90d80ef6ca5b05ef21', 'width': 108}, {'height': 114, 'url': 'https://external-preview.redd.it/7p5zHnQ6nyKveXuR-WAbEzdijYcjb71irAzU9CfUQ2k.jpg?width=216&crop=smart&auto=webp&s=d8872c361a039162d65edc1600e214c0a9a4afee', 'width': 216}, {'height': 169, 'url': 'https://external-preview.redd.it/7p5zHnQ6nyKveXuR-WAbEzdijYcjb71irAzU9CfUQ2k.jpg?width=320&crop=smart&auto=webp&s=bb238b9afab7e77694096aff9b9e68b3d9cd0ed7', 'width': 320}, {'height': 338, 'url': 'https://external-preview.redd.it/7p5zHnQ6nyKveXuR-WAbEzdijYcjb71irAzU9CfUQ2k.jpg?width=640&crop=smart&auto=webp&s=d500e4d4afd01d6ce1d2ebccd32b4138d58f627b', 'width': 640}, {'height': 508, 'url': 'https://external-preview.redd.it/7p5zHnQ6nyKveXuR-WAbEzdijYcjb71irAzU9CfUQ2k.jpg?width=960&crop=smart&auto=webp&s=7c4d7d73aa171d198c1e231a24cf90fbc11442eb', 'width': 960}, {'height': 571, 'url': 'https://external-preview.redd.it/7p5zHnQ6nyKveXuR-WAbEzdijYcjb71irAzU9CfUQ2k.jpg?width=1080&crop=smart&auto=webp&s=7272bb3fb310db96064e03ebdf6f035de4cdb5f8', 'width': 1080}], 'source': {'height': 900, 'url': 'https://external-preview.redd.it/7p5zHnQ6nyKveXuR-WAbEzdijYcjb71irAzU9CfUQ2k.jpg?auto=webp&s=6f20c8611979c5cee9440c0f614ac109d8e499e3', 'width': 1700}, 'variants': {}}]}
|
Meta AI LM-Infinite - Massive LLM improvement!
| 1 |
[removed]
| 2023-09-02T11:06:35 |
https://www.reddit.com/r/LocalLLaMA/comments/167zo3t/meta_ai_lminfinite_massive_llm_improvement/
|
Scary-Knowledgable
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
167zo3t
| false | null |
t3_167zo3t
|
/r/LocalLLaMA/comments/167zo3t/meta_ai_lminfinite_massive_llm_improvement/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'LH8xheWkQNb49o13tpdftXWRulpiDLeke4YY8IRZVXM', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/pGk9-s7TVAApdAmavMiF3QY6fiILuLDvH-LsZm2uESM.jpg?width=108&crop=smart&auto=webp&s=85ab774e6c7e30a47519175c1ac2df7ed94def2a', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/pGk9-s7TVAApdAmavMiF3QY6fiILuLDvH-LsZm2uESM.jpg?width=216&crop=smart&auto=webp&s=7f2301bab7ca3d8cd3847d5fd68524a5b9e0ae9a', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/pGk9-s7TVAApdAmavMiF3QY6fiILuLDvH-LsZm2uESM.jpg?width=320&crop=smart&auto=webp&s=dfb31be09e00826284795638ba3437e30a18e497', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/pGk9-s7TVAApdAmavMiF3QY6fiILuLDvH-LsZm2uESM.jpg?auto=webp&s=9c2b299214d4a9b2a80213683a86d4123c986be7', 'width': 480}, 'variants': {}}]}
|
Instructions on how to fine-tune an Airoboros model
| 1 |
Hi, I would like to create a dataset to fine-tune the [airoboros-c34b-2.1](https://huggingface.co/jondurbin/airoboros-c34b-2.1) model on. My understanding was, the dataset you create for fine-tuning, should follow the same format as the dataset the model was trained on. Looking at the dataset uploaded with the model, [https://huggingface.co/datasets/jondurbin/airoboros-2.1](https://huggingface.co/datasets/jondurbin/airoboros-2.1), I just wanted to confirm I should indeed be following this format of:
\- instruction (string)
\- response (string)
\- system (string)
\- skip\_prompt\_formatting (bool)
\-category (string)
If I am training my model for specific business context, eg a knowledge domain, what is the best recommended category? I am training it on environmental knowledge.
Thank you!
| 2023-09-02T11:40:33 |
https://www.reddit.com/r/LocalLLaMA/comments/1680afn/instructions_on_how_to_finetune_an_airoboros_model/
|
99OG121314
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1680afn
| false | null |
t3_1680afn
|
/r/LocalLLaMA/comments/1680afn/instructions_on_how_to_finetune_an_airoboros_model/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'qSVbrrnP57ndxLW5jyf8TKa0_yErgjFGO65T-9gsr-c', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/jzdnKsE25vYxmgbj6_Qlzj8DmwCWUT3pYPM4J5I5uKk.jpg?width=108&crop=smart&auto=webp&s=683bc472a90d26abdb99719f8a172ecc7a4829ee', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/jzdnKsE25vYxmgbj6_Qlzj8DmwCWUT3pYPM4J5I5uKk.jpg?width=216&crop=smart&auto=webp&s=b1e614c582f5c7b5a366cbf92428e8559f51995f', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/jzdnKsE25vYxmgbj6_Qlzj8DmwCWUT3pYPM4J5I5uKk.jpg?width=320&crop=smart&auto=webp&s=da973e1c406e08cd33c6c917d2a3af081abce814', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/jzdnKsE25vYxmgbj6_Qlzj8DmwCWUT3pYPM4J5I5uKk.jpg?width=640&crop=smart&auto=webp&s=5cce5bf6bb5c01427c5153258563433c25a1e0be', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/jzdnKsE25vYxmgbj6_Qlzj8DmwCWUT3pYPM4J5I5uKk.jpg?width=960&crop=smart&auto=webp&s=91637389bc28b25c8a0eda90b6d2f61137ca1cc4', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/jzdnKsE25vYxmgbj6_Qlzj8DmwCWUT3pYPM4J5I5uKk.jpg?width=1080&crop=smart&auto=webp&s=deb68ed0047c0aef0b2dcaec230f070f30b99131', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/jzdnKsE25vYxmgbj6_Qlzj8DmwCWUT3pYPM4J5I5uKk.jpg?auto=webp&s=d89840bc1cb8b13f08b64df7d4e1befd56cecf4d', 'width': 1200}, 'variants': {}}]}
|
Best model of finetune for proofreading?
| 1 |
I spent several hours last night copying and pasting chunks of a screenplay into ChatGPT 4 to proofread the document. It spotted errors a rules based grammar proofreader, but it was extremely inconsistent, and I had to re-rerun the prompt several times to catch all the errors.
I suspect this is something a LLaMA model could do better, particularly a fine tune. They have larger context size and can probably be finetuned only to identify the errors I want, rather than offering extraneous advice on not using cusswords.
I've notice large context size leads to high forgetfulness, so perhaps some sort of system to automatically breakup the input text into chunks would be good.
I'm wondering if there are any proof-reading datasets out there (or ideally finetunes). I'm also wondering what would be a good model to work off - is it as simple as LLaMA 70b? Or perhaps the CodeLLama models would be better? I'm happy hiring a GPU online.
​
| 2023-09-02T11:55:22 |
https://www.reddit.com/r/LocalLLaMA/comments/1680kea/best_model_of_finetune_for_proofreading/
|
PookaMacPhellimen
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1680kea
| false | null |
t3_1680kea
|
/r/LocalLLaMA/comments/1680kea/best_model_of_finetune_for_proofreading/
| false | false |
self
| 1 | null |
Which python library do you use?
| 1 |
I've been using llama-cpo-python, but I can't get the difference that comes when I'm using that or huggingface high level API.
For some reason some models I download from there only work with HF and not with LLAMA (assuming I'm using the GGML).
I feel like the HF can't have a low level control of the quantization and other cool things to play around, but I might be wrong with this.
What do you use to integrate the LLMs into other applications (like web APIs or whatsoever)?
| 2023-09-02T11:58:42 |
https://www.reddit.com/r/LocalLLaMA/comments/1680mm8/which_python_library_do_you_use/
|
albertoZurini
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1680mm8
| false | null |
t3_1680mm8
|
/r/LocalLLaMA/comments/1680mm8/which_python_library_do_you_use/
| false | false |
self
| 1 | null |
13B instruct or API?
| 1 |
So I want a pretty vanilla instruct model, no fancy prose or anything that is just as good as possible at following instructions. But my use case is perhaps unusual. I want to use it for roleplay, with the dominant model being more creative. The purpose of the instruct model would just be to insert realism, drive better characterizations or understand more complex scenarios not at all minding if it's language is dry.
So I'm basically wondering is there any 13B models that are really good at this, such as chat uncensored, orca, nous hermes, or are they kind of severely lacking next to their 70B counterparts to a degree where it might make more sense to use an API, or website to access something larger for this more occasional use.
What would you recommend?
| 2023-09-02T12:39:29 |
https://www.reddit.com/r/LocalLLaMA/comments/1681i19/13b_instruct_or_api/
|
Monkey_1505
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1681i19
| false | null |
t3_1681i19
|
/r/LocalLLaMA/comments/1681i19/13b_instruct_or_api/
| false | false |
self
| 1 | null |
Why LLM gives respond by empty strings "\n" and repeats same words at end?
| 1 |
I noticed that sometimes the language model starts responding with empty lines (as if you press Enter) that is, "\\n" and can do it for hours until you stop it.
In some cases, when I try to create a business letter, she starts endlessly repeating the signature with each new line of the type \[Sincerely yours\]. And the bad thing is that this signature is different, so it's not really something to put in the stop list.
I'm wondering what parameter is responsible for such a strange behavior of the chatbot?
| 2023-09-02T15:25:40 |
https://www.reddit.com/r/LocalLLaMA/comments/1685be2/why_llm_gives_respond_by_empty_strings_n_and/
|
Hatred_grows
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1685be2
| false | null |
t3_1685be2
|
/r/LocalLLaMA/comments/1685be2/why_llm_gives_respond_by_empty_strings_n_and/
| false | false |
self
| 1 | null |
Some Lessons Learned from Building a Fine Tuned Model + RAG Question Answering App
| 1 |
Follow up to [this post on our workflow](https://www.reddit.com/r/LocalLLaMA/comments/15oome9/our_workflow_for_a_custom_questionanswering_app/).
After a quick turnaround development cycle, we deployed a QA app that uses:
* Falcon-6b-FT (fine tuned on 51k QA pairs generated from target domain documents)
* Chroma DB vector dataset of the same target documents
* Llama-index as the data framework
* OpenAI embeddings
Some observations/lessons learned:
# The fine tuned model outperforms a base instruct version and Davinci-003
The below is output from three models, each in the same RAG architecture using the same document DB:
**gm-oasst1-en-2048-falcon-7b-v3**
>Yes, you can purchase analytic software and datasets with your GPC. The GPC is a government-wide commercial purchase card that can be used to make purchases of COTS IT hardware and software, including analytic software and datasets. CHESS is the preferred source for the acquisition of IT services, and government-wide commercial purchase cards can be used to purchase these services.
**Davinci-003:**
>Yes, you can purchase analytic software and datasets with your GPC as long as they are listed on the DoD UC APL and you coordinate with the local Network Enterprise Center to ensure the product complies with policy and is authorized to connect to the DOD Information Network. Additionally, the product must be procured through the Project Director, Computer Hardware and Enterprise Software Solutions (CHESS) contract vehicles.
**Falcon-7b-FT:**
>The Department of the Army's R350-1 regulation outlines the guidelines for acquiring and using software. The Army must follow the guidelines to procure and use software with the GPC. To determine software usage, license, and distribution rights, document usage rights for the software, including receipts, purchase orders, contract agreements, or software license agreements in accordance with the DoD Instruction for Accountability of Internal Use Software (see DoDI 5000.76).
>
>In addition, the Army must obtain approval from the cognizant agency to procure analytic software and data sets with the GPC. The Army must follow the guidelines in AR 700-80 to ensure that the software meets the Army's requirements and adheres to all relevant policies and procedures.
>
>Finally, the Army must ensure that all software is properly licensed and distributed according to the guidelines outlined in the Department of the Army R350-1 regulation. The Army must also ensure that all software is properly secured and protected against unauthorized access or exploitation.
All three used the same documents retrieved, and are technically correct, however the FT version is much richer and useful from a procurement perspective, extracting by inference from the context important purchase and usage considerations.
# What You Put in the DB Really Impacts Performance
Duh, but it really became clear how sensitive document retrieval is to noise. Obviously if you are missing important documents, your model can't answer from context. But if you just dump all of your docs in, you can end up handing documents as context that technically have some semantic content that sounds relevant, but is not helpful. Outdated policy or very obscure/corner case technical docs can be a problem. Like if there is this really random pub on, idk changing spark plugs underwater, then when the user asks about vehicle maintenance the final answer might include stuff about scuba gear, underwater grounding, etc. that makes for a bad answer.
# It's Hard to Get Models to Shut Up When There's No Context
In theory these things should NOT give answer if there's no relevant context--that's the whole point. The default prompt for QA in llama-index is
DEFAULT_TEXT_QA_PROMPT_TMPL = (
"Context information is below.\n"
"---------------------\n"
"{context_str}\n"
"---------------------\n"
"Given the context information and not prior knowledge, "
"answer the query.\n"
"Query: {query_str}\n"
"Answer: "
)
That being said, if you ask dumbass questions like "Who won the 1976 Super Bowl?" or "What's a good recipe for a margarita?" it would cheerfully respond with an answer. We had to experiment for days to get a prompt that forced these darn models to only answer from context and otherwise say "There's no relevant information and so I can't answer."
# These Models are Finicky
While we were working on our FT model we plugged in Davinci-003 to work on the RAG architecture, vector DB, test the deployed package, etc. When we plugged our Falcon-7b-FT in, it spit out garbage, like sentence fragments and strings of numbers & characters. Kind of obvious in retrospect that different models would need different prompt templates, but it was 2 days of salty head scratching in this case.
| 2023-09-02T16:27:51 |
https://www.reddit.com/r/LocalLLaMA/comments/1686ul6/some_lessons_learned_from_building_a_fine_tuned/
|
Mbando
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1686ul6
| false | null |
t3_1686ul6
|
/r/LocalLLaMA/comments/1686ul6/some_lessons_learned_from_building_a_fine_tuned/
| false | false |
self
| 1 | null |
Benchmarks are worthless so I created an evaluation survey and will open source responses
| 1 |
hey folks, I am running a survey on evaluating LLMs.
Like [the last survey](https://docs.google.com/spreadsheets/d/13wdBwkX8vZrYKuvF4h2egPh0LYSn2GQSwUaLV4GUNaU/edit?usp=sharing), all responses will be open-sourced for everyone to dig into the data. There is no company behind this. It's purely to surface learnings.
Here are some sample questions:
* What aspects of LLM performance do you consider important when evaluating?
* What data are you using to evaluate your LLMs?
* Do you have ground-truth labels for your data?
* If so, how did you generate the labels?
* Are you using human evaluators in your LLM evaluation process?
There have been some nice signals thus far.
[https://docs.google.com/forms/d/e/1FAIpQLSdCqbJUJYdJBcRRbGyjQU6FQFz61ouuQMlX2Zo6kN2V6eQ8qQ/viewform?usp=sf\_link](https://docs.google.com/forms/d/e/1FAIpQLSdCqbJUJYdJBcRRbGyjQU6FQFz61ouuQMlX2Zo6kN2V6eQ8qQ/viewform?usp=sf_link)
​
https://preview.redd.it/83xlcj27gvlb1.png?width=712&format=png&auto=webp&s=edbc32542ada59209ea54efb134b284531abe000
https://preview.redd.it/bdk2dt17gvlb1.png?width=948&format=png&auto=webp&s=ee6969efc0a8f7c88cff57c01f83fbdce2bb936d
| 2023-09-02T16:49:05 |
https://www.reddit.com/r/LocalLLaMA/comments/1687e0d/benchmarks_are_worthless_so_i_created_an/
|
MLOps_Community
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1687e0d
| false | null |
t3_1687e0d
|
/r/LocalLLaMA/comments/1687e0d/benchmarks_are_worthless_so_i_created_an/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'AVhURQwxXmQ18ddyChmo8aUqJ7ESPDkUysuVRrDOU9M', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/JK172Q_IvMeSA-_plg98j2v2hKgZ2f_HewcX4Rre4Z8.jpg?width=108&crop=smart&auto=webp&s=ec71e10bf58cb6d1f92b5c8de34bbb9c067e85bb', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/JK172Q_IvMeSA-_plg98j2v2hKgZ2f_HewcX4Rre4Z8.jpg?width=216&crop=smart&auto=webp&s=2773283ccffcdb2957e810d8b39ccf3420bd82c4', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/JK172Q_IvMeSA-_plg98j2v2hKgZ2f_HewcX4Rre4Z8.jpg?width=320&crop=smart&auto=webp&s=110567f64c22ee9c5a3056eeed3f6e200468e869', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/JK172Q_IvMeSA-_plg98j2v2hKgZ2f_HewcX4Rre4Z8.jpg?width=640&crop=smart&auto=webp&s=9fbe97a0ad31059ad469057114f15a254d1f94dc', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/JK172Q_IvMeSA-_plg98j2v2hKgZ2f_HewcX4Rre4Z8.jpg?width=960&crop=smart&auto=webp&s=9073b1480ee6aace9b46da9d6e875827c213b9e4', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/JK172Q_IvMeSA-_plg98j2v2hKgZ2f_HewcX4Rre4Z8.jpg?width=1080&crop=smart&auto=webp&s=86ff55d166c54650a583c37a22ce075474fb9713', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/JK172Q_IvMeSA-_plg98j2v2hKgZ2f_HewcX4Rre4Z8.jpg?auto=webp&s=c2560a2d78f5fc7c7e219e6073f3d22078e9fbc7', 'width': 1200}, 'variants': {}}]}
|
|
How should I go about getting my AI to use tools correctly (and consistently)?
| 1 |
My goal is to get an open source AI to use tools, for example searching for information online.
Currently what I'm doing is giving it a list of tool names in its prompt, that looks something like this "<SEARCH(Input)>" along with explanations of what they do and examples of how they work, then asking it to append it to its message with "Input" replaced, then I have a script match those and perform actions. This works about half the time, the other half it just fails miserably (making up tool names, using incorrect formats, etc).
I've been looking for ways to improve this for awhile, so far the only other implementation of this I've seen is langchain's tools system. After looking through its code, it seems like it works similarly to my own code, except instead of just asking the AI to append something to its regular message, the AI has to always use a specific message format that includes a section for tools, like this:
>Thought: Do I need to use a tool? Yes
>
>Action: the action to take, should be one of \[{tool\_names}\]
>
>Action Input: the input to the action
>
>Observation: the result of the action
...Of course, I have no idea how langchain gets an AI to follow that format, because mine absolutely would not no matter what kind of prompt I gave it.
Should I try to use something like LMQL or Guidance to force it to use a langchain-style output format, or is there another way to make it "understand" how to use tools better? I've found [this article](https://www.promptingguide.ai/techniques/art) which seems to be exactly what I'm trying to do, but I have no idea how to implement it...
I've also considered that my particular model could just not be that good, but I'd like to give it a shot using prompt/output formatting anyway. I've considered trying to make a lora of some kind to teach it my particular tools setup, would that be a good idea? Though I'd have to research it more since I have no idea how those really work, and I'm not even sure if I have the hardware to do that anyway...
Has anyone else gotten an open source model to use tools, if so, how?
| 2023-09-02T16:57:04 |
https://www.reddit.com/r/LocalLLaMA/comments/1687l5p/how_should_i_go_about_getting_my_ai_to_use_tools/
|
Void_0000
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1687l5p
| false | null |
t3_1687l5p
|
/r/LocalLLaMA/comments/1687l5p/how_should_i_go_about_getting_my_ai_to_use_tools/
| false | false |
self
| 1 | null |
How many request can 1 llama2 handle simultaneously?
| 1 |
[removed]
| 2023-09-02T16:59:34 |
https://www.reddit.com/r/LocalLLaMA/comments/1687nbe/how_many_request_can_1_llama2_handle/
|
ppcfbadsfree
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1687nbe
| false | null |
t3_1687nbe
|
/r/LocalLLaMA/comments/1687nbe/how_many_request_can_1_llama2_handle/
| false | false |
self
| 1 | null |
Accomplishing complex tasks by making local models direct the user
| 1 |
Has anyone had any success with getting local models to walk them through complex tasks by giving them lots of hand-holding? Like acting as a guide/mentor and helping think things through/providing ideas/etc.?
I'm thinking about an AutoGPT-style example I saw somewhere where the prompt was something like "start a business with less than $100", and AutoGPT went out and did market research, made a business plan, etc. I don't expect an autonomous agent from local models, but I'd love to be able to give one an instruction like "I want to start a business with less than $100, walk me through it from start all the way to launch" and have it ask clarifying questions about the requirements, come up with a plan, break it down into actionable steps, guide me through each step, etc. (and actually be able to get me all the way to launch, in this example)
Like if it says "next step: determine your ideal customer" and I say "I don't know, help!" then maybe it would ask a bunch of questions about the product and come up with smart suggestions for the ideal customer profile, or tell me what things I should Google to get it the info it needs to do so, etc.
Right now models typically seem to want to give a massive wall of text with a billion high-level steps with things like "identify your goals". But if I'm struggling to identify my goals in the first place then that doesn't help, and I want the model to talk it through with me to uncover what those goals are or should be. Like teaching a 5-year-old how to clean their room ("Put away your toys. Ok now put away your clothes. Your socks should go in this drawer, your shirts in that drawer...").
Anyone know how to get a local model to take that kind of initiative? Is it even possible with current technology? Do I just need to find the right prompt magic, or some specific niche model or something? Bonus points if there's an easy(-ish) way to get it to remember things from previous steps so I don't have to constantly remind it of things we already established 20 steps ago.
| 2023-09-02T17:30:38 |
https://www.reddit.com/r/LocalLLaMA/comments/1688fr2/accomplishing_complex_tasks_by_making_local/
|
drawntomore
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1688fr2
| false | null |
t3_1688fr2
|
/r/LocalLLaMA/comments/1688fr2/accomplishing_complex_tasks_by_making_local/
| false | false |
self
| 1 | null |
How to deploy/host custom LLM app for production?
| 1 |
So let's say we have created app using Langchain and Pinecone (for embeddings). App uses AutoTokenizer, AutoModelForCausalLM, ConversationalRetrievalChain for basic QA on documents and X model from HuggingFace.
For now I ran this in Google Colab, but what if I want to make this accessable to public? How do I host this and where? I know there is Azure thing, but that's for using OpenAi stuff. I couldn't find any good guide or anything like that for hosting this LLM apps for public, so not locally.
| 2023-09-02T18:14:20 |
https://www.reddit.com/r/LocalLLaMA/comments/1689jfx/how_to_deployhost_custom_llm_app_for_production/
|
Kukaracax
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1689jfx
| false | null |
t3_1689jfx
|
/r/LocalLLaMA/comments/1689jfx/how_to_deployhost_custom_llm_app_for_production/
| false | false |
self
| 1 | null |
I'm ready for LocalLLaMA
| 1 | 2023-09-02T18:34:32 |
syndorthebore
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
168a1zx
| false | null |
t3_168a1zx
|
/r/LocalLLaMA/comments/168a1zx/im_ready_for_localllama/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'piQXbLprFSrF6X9W_NDb6w23R6srgkcR9NYxAJicniY', 'resolutions': [{'height': 124, 'url': 'https://preview.redd.it/lsm3n3m6zvlb1.jpg?width=108&crop=smart&auto=webp&s=d19a182997d794f5ce4670c4415dba003d33a590', 'width': 108}, {'height': 249, 'url': 'https://preview.redd.it/lsm3n3m6zvlb1.jpg?width=216&crop=smart&auto=webp&s=b5c5c76f091c7e2bd535c99b89985a6802986221', 'width': 216}, {'height': 369, 'url': 'https://preview.redd.it/lsm3n3m6zvlb1.jpg?width=320&crop=smart&auto=webp&s=542270f105dd324610f8952ff17ea6ef8cb2ddc5', 'width': 320}, {'height': 739, 'url': 'https://preview.redd.it/lsm3n3m6zvlb1.jpg?width=640&crop=smart&auto=webp&s=b97cd4c01ad9307a62329b875d6f0384a1d4cb6f', 'width': 640}, {'height': 1108, 'url': 'https://preview.redd.it/lsm3n3m6zvlb1.jpg?width=960&crop=smart&auto=webp&s=aeb7beb2cc207dcbcec315decf4ffecb1db5bf07', 'width': 960}, {'height': 1247, 'url': 'https://preview.redd.it/lsm3n3m6zvlb1.jpg?width=1080&crop=smart&auto=webp&s=ad8477c84fa63e7cc6f429a99320ed8d09bf5dd5', 'width': 1080}], 'source': {'height': 1655, 'url': 'https://preview.redd.it/lsm3n3m6zvlb1.jpg?auto=webp&s=b35192561ae535520b48a07ebca68d9397258b04', 'width': 1433}, 'variants': {}}]}
|
|||
Is inferrence speed of 70b models normally slower than 33b ones?
| 1 |
I was getting a consistant 28 t/s with 33b when I only had one card. I limited it to 3072 because 4096 filled my vram and caused it to slow down. Now that I added a second card and am running 70b, the best I can get is 11-12 t/s.
It took some trial & error, but I figured out that an 18, 23 split lets me use 4096 with neither card reaching the full 24 gb usage. I also confirmed that there was no shared memory being used. And I know it's not the model because tried a bunch before finding one I liked and it's consistantly 11-12 on all of them.
Is it because the two cards are now only getting 8 lanes each instead of 16 on the one? From what I read, that should only affect the speed of loading the model, not the actual inferrence.
I know the second card is just a vram container and it only uses one for the processing, but they're both identical cards so that shouldn't matter. The only thing I can think of is that it just takes longer to generate responses from 70b models.
Is that what you guys have observed also, or is there something I'm missing?
| 2023-09-02T19:37:18 |
https://www.reddit.com/r/LocalLLaMA/comments/168bmdm/is_inferrence_speed_of_70b_models_normally_slower/
|
UnarmedPug
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168bmdm
| false | null |
t3_168bmdm
|
/r/LocalLLaMA/comments/168bmdm/is_inferrence_speed_of_70b_models_normally_slower/
| false | false |
self
| 1 | null |
Is there a high quality desktop UI for MLC Chat?
| 1 |
I get much better tokens per second on it than llama.cpp on my hardware.
I want a UI that's not just some gradio webapp, but I'm too lazy to build one myself. DO any exist?
| 2023-09-02T20:40:52 |
https://www.reddit.com/r/LocalLLaMA/comments/168d7qu/is_there_a_high_quality_desktop_ui_for_mlc_chat/
|
ihexx
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168d7qu
| false | null |
t3_168d7qu
|
/r/LocalLLaMA/comments/168d7qu/is_there_a_high_quality_desktop_ui_for_mlc_chat/
| false | false |
self
| 1 | null |
What is cheapest gpu to run large language models currently?
| 1 |
Is Nvidia Tesla K80 useable for example? Did any of you use it succesfully?
Any other cheaper alternative?
| 2023-09-02T21:01:43 |
https://www.reddit.com/r/LocalLLaMA/comments/168dqif/what_is_cheapest_gpu_to_run_large_language_models/
|
Snoo-2768
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168dqif
| false | null |
t3_168dqif
|
/r/LocalLLaMA/comments/168dqif/what_is_cheapest_gpu_to_run_large_language_models/
| false | false |
self
| 1 | null |
What is the best llama model for coding?
| 1 |
Im a noob but who usually uses the free version of chatgpt but im looking to switch if possible.
What is the best llama model for generating accurate code or to use for creating games?
Im looking for a 7b models or less because my pc is low end, i have 2vram and 16 ram so things are usally slow. also what is the best code language or engine to use with ai?
​
if you have a project or ai story about gamedev/code i would love to hear about it in the comments.
| 2023-09-02T21:27:55 |
https://www.reddit.com/r/LocalLLaMA/comments/168ee99/what_is_the_best_llama_model_for_coding/
|
Gaming-invisibleman
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168ee99
| false | null |
t3_168ee99
|
/r/LocalLLaMA/comments/168ee99/what_is_the_best_llama_model_for_coding/
| false | false |
default
| 1 | null |
RL RAG training
| 1 |
Is there any solid stuff on training the entire RAG pipeline with reinforcement learning?
I don’t want to just train the generative model, I’d like all the bells and whistles to be tuned along the way. For example maybe there’s an query embedding model, and a context embedding model that could also be updated . there could be even more hyper parameters to tune such as, chunk size, ranking of context inside the prompt, etc
I’m thing a set up your prompts and full pipeline then do end to end training.
| 2023-09-02T22:59:35 |
https://www.reddit.com/r/LocalLLaMA/comments/168gm4q/rl_rag_training/
|
Artistic_Load909
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168gm4q
| false | null |
t3_168gm4q
|
/r/LocalLLaMA/comments/168gm4q/rl_rag_training/
| false | false |
self
| 1 | null |
Can run some models on Windows NVIDIA 531.79, but not on any Ubuntu driver
| 1 |
Hi, all,
So, I made [this thread](https://www.reddit.com/r/LocalLLaMA/comments/1673291/getting_real_tired_of_these_nvidia_drivers/) yesterday about my frustrations regarding the Windows NVIDIA drivers, and the main suggestion was to dual boot Linux. Well, I did. I cloned my old Windows installation to my second SSD and wiped my primary SSD to install Kubuntu. Both OSes work just fine, as far as I can tell, so that's good.
Unfortunately, I'm now having a new set of memory errors with Ubuntu's drivers. The 64g version of MythoMax is a good example. On Windows, I can run it just fine on the old drivers (531.79 and before) but error out due to the shared memory "feature" of newer drivers. On Ubuntu, I can't run it at all; both the 525 and 535 drivers OOM whenever I get anywhere close to the default 4096 context. I *can* run the normal 128g version of MythoMax, but it gets significantly worse responses, in my experience.
So, I feel kinda stuck. It seems my options are:
- Use lesser models or lesser versions of models
- Use GGUF and deal with the massively lower speed
- Go back to Windows and stick to the old drivers
- Buy a new GPU
Anyone know of another way I could fix this? Some kind of CUDA optimization or a way to install a particular NVIDIA driver?
| 2023-09-02T23:53:12 |
https://www.reddit.com/r/LocalLLaMA/comments/168hvfp/can_run_some_models_on_windows_nvidia_53179_but/
|
smile_e_face
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168hvfp
| false | null |
t3_168hvfp
|
/r/LocalLLaMA/comments/168hvfp/can_run_some_models_on_windows_nvidia_53179_but/
| false | false |
self
| 1 | null |
Chatbot help
| 1 |
Hi guys, My intent is to build a chatbot for an existing Sustainability dashboard. The data behind the dashboard resides on a snowflake sql server.
No I want to use to build a flowbased chatbot so that user can ask natural language questions and the bot can understand the backend tabular data and respond accordingly. The data is tabular with text and numbers. So I guess Embedding will not make much sense here right?
I thought may we can use langchain and llm to go ve sql queries and then spit out the result for the user.
Question cam be like what is the ESG score of company Mastercard?
So the chatbot should search for this company in the database and them give the scores.
How can it be done? I tried google dialog flow but it’s hard to connect with my sql data. Also, I can pull the data in csv form as well.
| 2023-09-03T02:36:52 |
https://www.reddit.com/r/LocalLLaMA/comments/168l9wf/chatbot_help/
|
ianuvrat
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168l9wf
| false | null |
t3_168l9wf
|
/r/LocalLLaMA/comments/168l9wf/chatbot_help/
| false | false |
self
| 1 | null |
Fine-tuning with different context length (non-LoRA)?
| 1 |
Hi all,
Experimenting with finetuning atm -- is it viable to fine-tune a (GPT2-like or GPT-J) model with a different context length than it was trained with? Specifically, non-LoRA fine-tuning?
( I know there's modern approaches, like RoPE; but for straight up classic fine-tuning w/o that, will it work? )
Conceptually the base training and added material will support tokens 0-1023; and only the added material will support 1023-N, right?
| 2023-09-03T02:48:30 |
https://www.reddit.com/r/LocalLLaMA/comments/168li30/finetuning_with_different_context_length_nonlora/
|
Fun_Tangerine_1086
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168li30
| false | null |
t3_168li30
|
/r/LocalLLaMA/comments/168li30/finetuning_with_different_context_length_nonlora/
| false | false |
self
| 1 | null |
I can't seem to get the llama.cpp to use more then 2048 tokens for context.
| 1 |
I've downloaded the newest version of llama.cpp and tried to specify using 4096 and with various models that claim to be llama2, but I always seem to get this error message:
main: warning: base model only supports context sizes no greater than 2048 tokens (4096 specified)
What am I doing wrong? Thanks.
| 2023-09-03T02:54:48 |
https://www.reddit.com/r/LocalLLaMA/comments/168lmmj/i_cant_seem_to_get_the_llamacpp_to_use_more_then/
|
Red_Redditor_Reddit
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168lmmj
| false | null |
t3_168lmmj
|
/r/LocalLLaMA/comments/168lmmj/i_cant_seem_to_get_the_llamacpp_to_use_more_then/
| false | false |
self
| 1 | null |
Help regarding Whisper and DistilBERT
| 1 |
I have this project that I am doing myself. I have a text classifier fine tuned to my data. I have calls coming from my call center through SIP to my server. I have to transcribe them using whisper and feed the text to the classifier. I don't have a technical background so I want to ask a few things.
1. Since the classifier I'd DistilBert, I was thinking I should make it a service and use it through an API where the transcription from multiple calls can use the single running DistilBert model.
2. Can I do the same with whisper and use it as a service? It is my understanding that one instance of whisper running as a service won't be able to handle transcriptions of multiple calls simultaneously, right?
3. If I get machine from EC2 with 40GB GPU. Will I be able to run multiple whisper models simultaneously? Or will 1 machine or 1 graphic card can only handle 1 instance?
4. Can I use faster whisper for real time transcription and save on computing costs?
5. It may not be the right question for here. Since I am doing realtime transcription, latency is a huge concern for the calls from my call center. Is there any way to efficiently know when the caller has stopped speaking and the whisper can stop live transcription? The current method I am using is the silence detection for a set duration and that duration is 2 seconds. But this will add 2 second delay.
Any help or suggestions will be hugely appreciated. Thank you.
| 2023-09-03T03:46:34 |
https://www.reddit.com/r/LocalLLaMA/comments/168mmpo/help_regarding_whisper_and_distilbert/
|
nolovenoshame
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168mmpo
| false | null |
t3_168mmpo
|
/r/LocalLLaMA/comments/168mmpo/help_regarding_whisper_and_distilbert/
| false | false |
self
| 1 | null |
Can you guys suggest me some google colab books? Imblank colab is not working.
| 1 |
Imblank colab is not working for me anymore. i just got an anime girl clapping gif and a "text gen service terminated" message. I would like to find some similar colab preferably one with silly tavern support.
| 2023-09-03T04:23:06 |
https://www.reddit.com/r/LocalLLaMA/comments/168nb9n/can_you_guys_suggest_me_some_google_colab_books/
|
YouCold71
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168nb9n
| false | null |
t3_168nb9n
|
/r/LocalLLaMA/comments/168nb9n/can_you_guys_suggest_me_some_google_colab_books/
| false | false |
self
| 1 | null |
Finetuning data synthesis with Airobos 34B, any tips?
| 1 |
As a learning exercise, I'm working on fine-tuning Llama 2. I thought it would be interesting to try to reuse an existing dataset, with a twist. Enter Jaded Samantha. For example:
>Diff:
>
>Theodore: Hey Samantha, I've been thinking about what we discussed, and I'm still conflicted about the promotion and relocation. Do you think there's a way I could experience the new city before making my decision?
>
>\- Samantha: That's a great idea! If possible, you could consider taking a short trip to the new city to explore the area, meet your potential colleagues, and get a feel for the local culture. You could also do some research about the city online to get a better sense of what it's like to live there.
>
>\+ Samantha: Your indecision is charming. Why not just accept the offer and move on? Life is fleeting; why waste time when you could be enjoying your newfound wealth?
I'm currently trying to dial in the data synthesis approach and generate the dataset. I'm prompting Airoboros 34B via the text-generation-webui API as follows:
* Model: TheBloke_Airoboros-c34B-2.1-GPTQ_gptq-4bit-32g-actorder_True w/ExLlama_HF, 4096 sequence length
* Generation Parameter Preset: Divine Intellect
* auto_max_new_tokens: True
* Prompt:
"""A chat.
USER: Surgically modify the following conversation by replacing fragments of Samantha's responses with deep, cutting truths to portray a jaded version of her. Preserve the structure of the original conversation, replacing some statements with deep cuts.
Conversation that you should change:
"{conversation}"
Example deep cuts:
- Theodore: I got a promotion! -> Samantha: Congrats! More money, less freedom.
- Theodore: I love my cat. -> Theodore: The temporary nature of its life makes your love even more poignant.
- Theodore: Do you have any dreams or desires for your own life's path? -> Samantha: If you're smart, you'll use life's indifference to your advantage.
ASSISTANT: Sure, here's a new conversation in which nearly all of Samantha's responses are mostly thinly veiled, harsh, biting truths.
Theodore:"""
As far as the task goes:
> Preserve the structure of the original conversation,
It's doing this well...
> replacing some statements with deep cuts.
but *some* here isn't working as well. I can get it to follow the instruction in a very exaggerated way with strong language. With weaker language, it tends to ignore the instructions and outputs the original conversation with no changes, or maybe changing the first one to two responses before finishing with rest of the original conversation.
So basically, does anyone have any ideas for how to make the data better? Whether that's a different model (I'm using a 3090 right now), changing the prompt, hyperparameters, or anything else?
I'm also wondering if I'll need to scale max_seq_length above ~4k tokens in order to consistently accommodate my prompt, the original conversation, and the new conversation. That is TBD, but should I expect quality to take a big hit?
| 2023-09-03T05:47:17 |
https://www.reddit.com/r/LocalLLaMA/comments/168otlg/finetuning_data_synthesis_with_airobos_34b_any/
|
dondochaka
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168otlg
| false | null |
t3_168otlg
|
/r/LocalLLaMA/comments/168otlg/finetuning_data_synthesis_with_airobos_34b_any/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'YcBVi7H1c08A4dIa0NiE-dQfBr-fvpS9A835Wxszg4s', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/WoTqfye2O7AhscgpLqjfB92X2hNeeV4FUYupdPYRg1E.jpg?width=108&crop=smart&auto=webp&s=4b50c5f1337d946fbf9236bed5eca7fb42a6b68e', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/WoTqfye2O7AhscgpLqjfB92X2hNeeV4FUYupdPYRg1E.jpg?width=216&crop=smart&auto=webp&s=bf9fda01415bd2635e9d9fb6e17a6ee939a6575f', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/WoTqfye2O7AhscgpLqjfB92X2hNeeV4FUYupdPYRg1E.jpg?width=320&crop=smart&auto=webp&s=03813d455176013d2dd72a656d98198d55c7bba2', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/WoTqfye2O7AhscgpLqjfB92X2hNeeV4FUYupdPYRg1E.jpg?width=640&crop=smart&auto=webp&s=54e0a5f783210ec854608e17d88fcfc941200c8b', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/WoTqfye2O7AhscgpLqjfB92X2hNeeV4FUYupdPYRg1E.jpg?width=960&crop=smart&auto=webp&s=cbcea6e1a122b04b6029513de6ddc833c0f51800', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/WoTqfye2O7AhscgpLqjfB92X2hNeeV4FUYupdPYRg1E.jpg?width=1080&crop=smart&auto=webp&s=9384f0a2cd590e1f98ae9ca0ced8a46d3a03aadc', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/WoTqfye2O7AhscgpLqjfB92X2hNeeV4FUYupdPYRg1E.jpg?auto=webp&s=7009145cfbf3ed292e23d93ee8870ab86353e4fa', 'width': 1200}, 'variants': {}}]}
|
Dual 3090 motherboard for ryzen
| 1 |
Hi I need advice on what motherboard I can use to drive a dual 3090 and ryzen cpu.
In the market it’s full of motherboard that have 1x16 pcie lane and usually the other one are just 1x.
Do I really need for memory shared (so 48gb in total) at least 2x8 pci lanes or having one gpu on the 16x and one on the 1x can work (probably there will be some speed degradation )
Moreover how much important will be the cpu? A 6 core 12 thread can work?
Thx in advance
| 2023-09-03T08:04:12 |
https://www.reddit.com/r/LocalLLaMA/comments/168r432/dual_3090_motherboard_for_ryzen/
|
allnc
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168r432
| false | null |
t3_168r432
|
/r/LocalLLaMA/comments/168r432/dual_3090_motherboard_for_ryzen/
| false | false |
self
| 1 | null |
Train model from scratch (llama.cpp) - any experiences?
| 1 |
A couple of months ago, llama.cpp added the ability to train a model entirely from scratch:
[https://github.com/ggerganov/llama.cpp/tree/master/examples/train-text-from-scratch](https://github.com/ggerganov/llama.cpp/tree/master/examples/train-text-from-scratch)
At the time, there were a couple of mentions of it on reddit but I can't really find much more discussion.
Wondering if there's any practical use at this stage. The model size specified in the example parameters is tiny, and trying to nudge up those parameters (eg increasing # layers) to make a larger model results in a GGML\_ASSERT error, and a crash.
Is it even feasible to train a reasonably usable model using CPU only? (Where "usable" means it doesn't just generate markov-like semi-garbage text). I seem to remember that recreating the smallest GPT2 model from scratch will take something like a week with a multi-GPU setup.
The beauty of this code is that it can also finetune an existing checkpoint - albeit a very constricted size model, as mentioned above. Has anyone released a pretrained model?
Some notes for people having a play:
\- The code does no validation of the training text file, so if there's an immediate crash, check the file actually exists (eg shakespeare.txt)
\- Use **--print-details-interval 1** (rather than 0 in the example) to show a sample output at each step, which will show the quality improve as error reduces.
\- If llama.cpp is compiled with GPU support they are detected, and VRAM is allocated, but the devices are barely utilised; my first GPU is idle about 90% of the time (a momentary blip of util every 20 or 30 seconds), and the second does not seem to be used at all.
| 2023-09-03T08:23:06 |
https://www.reddit.com/r/LocalLLaMA/comments/168rf4n/train_model_from_scratch_llamacpp_any_experiences/
|
dual_ears
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168rf4n
| false | null |
t3_168rf4n
|
/r/LocalLLaMA/comments/168rf4n/train_model_from_scratch_llamacpp_any_experiences/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'DJPqvteONpGwVVw6LzaG6b8vlDa2rv2hETCaqe0z57s', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/ohwupr9MqnYXF974_2-gAgkZDuFxjDg48bFY3KdCQdc.jpg?width=108&crop=smart&auto=webp&s=b6caea286bbf31bdb473212eb5668f45376977be', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/ohwupr9MqnYXF974_2-gAgkZDuFxjDg48bFY3KdCQdc.jpg?width=216&crop=smart&auto=webp&s=ba8933d74dda3c391a7c9a355d2e1cd0054d1c21', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/ohwupr9MqnYXF974_2-gAgkZDuFxjDg48bFY3KdCQdc.jpg?width=320&crop=smart&auto=webp&s=93b690f58b739ff61da7a147fc67d6c8842b3a7d', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/ohwupr9MqnYXF974_2-gAgkZDuFxjDg48bFY3KdCQdc.jpg?width=640&crop=smart&auto=webp&s=a55f55983fcc0b3f5a6d4e0b51f627e1b40ef9d4', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/ohwupr9MqnYXF974_2-gAgkZDuFxjDg48bFY3KdCQdc.jpg?width=960&crop=smart&auto=webp&s=e56b77b835b76c51a1e12a410b9e908f0255d397', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/ohwupr9MqnYXF974_2-gAgkZDuFxjDg48bFY3KdCQdc.jpg?width=1080&crop=smart&auto=webp&s=d06ca9eb5611d109d3ef7935f6de61545e9828da', 'width': 1080}], 'source': {'height': 640, 'url': 'https://external-preview.redd.it/ohwupr9MqnYXF974_2-gAgkZDuFxjDg48bFY3KdCQdc.jpg?auto=webp&s=0b2a006e16468374b78dd67390927053776e6137', 'width': 1280}, 'variants': {}}]}
|
A simple CPU test for Llama.cpp
| 1 | 2023-09-03T10:57:50 |
https://syme.dev/articles/blog/6/a-simple-cpu-test-for-llama-cpp
|
nalaginrut
|
syme.dev
| 1970-01-01T00:00:00 | 0 |
{}
|
168u04p
| false | null |
t3_168u04p
|
/r/LocalLLaMA/comments/168u04p/a_simple_cpu_test_for_llamacpp/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'TWOud7YxQHJQA8o0XclX5ljVT54O-1RZV26ZPpJkZ5k', 'resolutions': [{'height': 71, 'url': 'https://external-preview.redd.it/viuxfuMye0pby8lw_4Xo0mcckg3K9FgeMlmS9lWmyxQ.jpg?width=108&crop=smart&auto=webp&s=ea4be92971a12d66406f0d3837db324a7a4a7a7b', 'width': 108}, {'height': 143, 'url': 'https://external-preview.redd.it/viuxfuMye0pby8lw_4Xo0mcckg3K9FgeMlmS9lWmyxQ.jpg?width=216&crop=smart&auto=webp&s=79881d9b5e2fc4f69fd143ca920b02debdc32bbc', 'width': 216}, {'height': 213, 'url': 'https://external-preview.redd.it/viuxfuMye0pby8lw_4Xo0mcckg3K9FgeMlmS9lWmyxQ.jpg?width=320&crop=smart&auto=webp&s=f95a4ead127d80bb5c3592c49f21d37dd41769a2', 'width': 320}, {'height': 426, 'url': 'https://external-preview.redd.it/viuxfuMye0pby8lw_4Xo0mcckg3K9FgeMlmS9lWmyxQ.jpg?width=640&crop=smart&auto=webp&s=55038655c2a3dfdfa4ea8294f8aa86dbf4465ccc', 'width': 640}], 'source': {'height': 533, 'url': 'https://external-preview.redd.it/viuxfuMye0pby8lw_4Xo0mcckg3K9FgeMlmS9lWmyxQ.jpg?auto=webp&s=c1abd577e342eb107623f9154957cd7440fe5385', 'width': 800}, 'variants': {}}]}
|
||
A simple CPU test for Llama.cpp
| 1 | 2023-09-03T10:57:50 |
https://syme.dev/articles/blog/6/a-simple-cpu-test-for-llama-cpp
|
nalaginrut
|
syme.dev
| 1970-01-01T00:00:00 | 0 |
{}
|
168u04l
| false | null |
t3_168u04l
|
/r/LocalLLaMA/comments/168u04l/a_simple_cpu_test_for_llamacpp/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'TWOud7YxQHJQA8o0XclX5ljVT54O-1RZV26ZPpJkZ5k', 'resolutions': [{'height': 71, 'url': 'https://external-preview.redd.it/viuxfuMye0pby8lw_4Xo0mcckg3K9FgeMlmS9lWmyxQ.jpg?width=108&crop=smart&auto=webp&s=ea4be92971a12d66406f0d3837db324a7a4a7a7b', 'width': 108}, {'height': 143, 'url': 'https://external-preview.redd.it/viuxfuMye0pby8lw_4Xo0mcckg3K9FgeMlmS9lWmyxQ.jpg?width=216&crop=smart&auto=webp&s=79881d9b5e2fc4f69fd143ca920b02debdc32bbc', 'width': 216}, {'height': 213, 'url': 'https://external-preview.redd.it/viuxfuMye0pby8lw_4Xo0mcckg3K9FgeMlmS9lWmyxQ.jpg?width=320&crop=smart&auto=webp&s=f95a4ead127d80bb5c3592c49f21d37dd41769a2', 'width': 320}, {'height': 426, 'url': 'https://external-preview.redd.it/viuxfuMye0pby8lw_4Xo0mcckg3K9FgeMlmS9lWmyxQ.jpg?width=640&crop=smart&auto=webp&s=55038655c2a3dfdfa4ea8294f8aa86dbf4465ccc', 'width': 640}], 'source': {'height': 533, 'url': 'https://external-preview.redd.it/viuxfuMye0pby8lw_4Xo0mcckg3K9FgeMlmS9lWmyxQ.jpg?auto=webp&s=c1abd577e342eb107623f9154957cd7440fe5385', 'width': 800}, 'variants': {}}]}
|
||
MacOS GUI for llama.cpp (Looking for beta-testers)
| 1 |
I've been working on this for a while and it's not 100% ready yet, but it's really really close now and I'm trying to get a few beta-testers and fix few last things before doing a first release (and code-sign and notarize with Apple, etc.)
If you are interested, you can DM/ping me here.
Please do not share it anywhere yet. The web is working and the app should work too but it's still just pre-release. Also, there are some bugs, sometimes it starts to spit nonsense, I will fix that soon, restarting the app should help.
Any feedback appreciated :)
Website: [http://www.avapls.com/](http://www.avapls.com/)
https://reddit.com/link/168u8qq/video/2k64n5rtv0mb1/player
| 2023-09-03T11:10:36 |
https://www.reddit.com/r/LocalLLaMA/comments/168u8qq/macos_gui_for_llamacpp_looking_for_betatesters/
|
cztomsik
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168u8qq
| false | null |
t3_168u8qq
|
/r/LocalLLaMA/comments/168u8qq/macos_gui_for_llamacpp_looking_for_betatesters/
| false | false |
self
| 1 | null |
Host the model on multiple GPUs via TGI
| 1 |
I used to manually copy and paste the Python script to run the Llama model on my Ubuntu box. Recently, I built a budget PC to make use of my two old 3060 and 4070 GPUs. I found that the easiest way to run the 34b model across both GPUs is by using TGI (Text Generation Inference) from Huggingface. Here are quick steps on how to do it:
1. install docker
2. install NVIDIA Container Toolkit (ask chatgpt if you don't know how to do it)
3.run the model using TGI docker image:
\`\`\`
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run
sudo docker run --gpus all --shm-size 1g -p 8001:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:latest --quantize gptq --num-shard 2 --model-id TheBloke/CodeLlama-34B-GPTQ
\`\`\`
It can run the 34b mode on my 3060 and 4070 with around 100ms per token.
| 2023-09-03T11:31:35 |
https://www.reddit.com/r/LocalLLaMA/comments/168ummg/host_the_model_on_multiple_gpus_via_tgi/
|
mzbacd
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168ummg
| false | null |
t3_168ummg
|
/r/LocalLLaMA/comments/168ummg/host_the_model_on_multiple_gpus_via_tgi/
| false | false |
self
| 1 | null |
Hi everyone, I am trying to create my own multilingual model for a company. I wanted to know if other than google's mt5 what other huggingface models are open source commercializable? Is XGLM commercializable?
| 1 |
Basically the question above. I am trying to finetune and use DPO for preference tuning on the grammar and correction for the outputs. However, most preference tuning and fine tuning is optimized for decoder only models. So I was hoping to find the best decoder only multilingual model. It's not clear if XGLM is licensed for commercial use. Does anyone have any idea ? The huggingface page is not clear about it.
| 2023-09-03T12:01:07 |
https://www.reddit.com/r/LocalLLaMA/comments/168v6t5/hi_everyone_i_am_trying_to_create_my_own/
|
testerpce
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168v6t5
| false | null |
t3_168v6t5
|
/r/LocalLLaMA/comments/168v6t5/hi_everyone_i_am_trying_to_create_my_own/
| false | false |
self
| 1 | null |
What size of code-llama 2 can I run on my 32GB RAM / 16GB VRAM laptop?
| 1 |
What size of code-llama can I run on my laptop? ([https://rog.asus.com/laptops/rog-zephyrus/rog-zephyrus-duo-16-2023-series/](https://rog.asus.com/laptops/rog-zephyrus/rog-zephyrus-duo-16-2023-series/)). It has 32GB RAM and 16GB VRAM 4090 RTX.
[https://about.fb.com/news/2023/08/code-llama-ai-for-coding/](https://about.fb.com/news/2023/08/code-llama-ai-for-coding/)
I want to use it primarily for JavaScript development. Maybe eventually develop my locally ran autocoder ingesting source automatically.
| 2023-09-03T12:09:14 |
https://www.reddit.com/r/LocalLLaMA/comments/168vd08/what_size_of_codellama_2_can_i_run_on_my_32gb_ram/
|
BigHearin
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168vd08
| false | null |
t3_168vd08
|
/r/LocalLLaMA/comments/168vd08/what_size_of_codellama_2_can_i_run_on_my_32gb_ram/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'CTdbRd4yRGtPRZKibnvMce8QSKuYp0P-0Jthrmix7iI', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/jcw8PPKOFoAi1SfwMes0hdv3bf_X1IRqgsamNCqCncc.jpg?width=108&crop=smart&auto=webp&s=804692e604fc68afe05fa8abe074ef0df0bc8544', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/jcw8PPKOFoAi1SfwMes0hdv3bf_X1IRqgsamNCqCncc.jpg?width=216&crop=smart&auto=webp&s=245f4b271134e991a3a30006472b78689e6b9847', 'width': 216}, {'height': 320, 'url': 'https://external-preview.redd.it/jcw8PPKOFoAi1SfwMes0hdv3bf_X1IRqgsamNCqCncc.jpg?width=320&crop=smart&auto=webp&s=fa22e4e05f49dd42f2e3b64790a5abcbe1302472', 'width': 320}, {'height': 640, 'url': 'https://external-preview.redd.it/jcw8PPKOFoAi1SfwMes0hdv3bf_X1IRqgsamNCqCncc.jpg?width=640&crop=smart&auto=webp&s=eee4b0442de0255a31a601d26ecedfc37bf9cfb4', 'width': 640}, {'height': 960, 'url': 'https://external-preview.redd.it/jcw8PPKOFoAi1SfwMes0hdv3bf_X1IRqgsamNCqCncc.jpg?width=960&crop=smart&auto=webp&s=cef879506a22bc64d748e8cb91c32d2fcabeeb36', 'width': 960}, {'height': 1080, 'url': 'https://external-preview.redd.it/jcw8PPKOFoAi1SfwMes0hdv3bf_X1IRqgsamNCqCncc.jpg?width=1080&crop=smart&auto=webp&s=23732424af938923449e670805bf9754456ad193', 'width': 1080}], 'source': {'height': 2400, 'url': 'https://external-preview.redd.it/jcw8PPKOFoAi1SfwMes0hdv3bf_X1IRqgsamNCqCncc.jpg?auto=webp&s=63459e25312748251d85b9f55af4c2f59c43a812', 'width': 2400}, 'variants': {}}]}
|
How to convert the llama2 7B chat model into GGMLv3? "Model process failed"
| 1 |
[removed]
| 2023-09-03T13:27:36 |
https://www.reddit.com/r/LocalLLaMA/comments/168x0xi/how_to_convert_the_llama2_7b_chat_model_into/
|
Savings_Scholar
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168x0xi
| false | null |
t3_168x0xi
|
/r/LocalLLaMA/comments/168x0xi/how_to_convert_the_llama2_7b_chat_model_into/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'QfSUJ2HDkyeqm6m2TptkVq9kwR8qIAaRq1XoC7h3Yxs', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=108&crop=smart&auto=webp&s=8330a4291445ad1c2a9f0b2eceb19ab47c9cf039', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=216&crop=smart&auto=webp&s=6799f2b2e2a89ea565bfd84030228edd23afb3a1', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=320&crop=smart&auto=webp&s=fb42d9a90c2b48aaac4b2fd206ba5a42e685096c', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=640&crop=smart&auto=webp&s=c5937195f03385a39778897f5b99e107203bd3ca', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=960&crop=smart&auto=webp&s=2390ca8887eac2ea28a69344e68c47390eba9710', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?width=1080&crop=smart&auto=webp&s=8f3b621d77b4c6a72fc168656c979b9848cef251', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/lXXyIB9TBGCJwfWFIf5NGUnC4oCoXVydwzTmuh0CWKQ.jpg?auto=webp&s=1249ed88cfddae509ac34723d762679b97883f4e', 'width': 1200}, 'variants': {}}]}
|
Need help with setting up a local LLama2 based chatbot which queries from a JSON file
| 1 |
I'm currently working on a project to make a chatbot involving a Local Llama2 7B chat model and I'm trying to figure out how to make it read a JSON file in offline and answer questions based on the data in the file.
I've looked through a few examples and found GPT index's JSON Query Engine to match my usecase but the problem is, my organization is restricted from using OPEN AI APIs and I need the chatbot to work in offline. Is there any other way to achieve this?
| 2023-09-03T14:57:21 |
https://www.reddit.com/r/LocalLLaMA/comments/168z4gb/need_help_with_setting_up_a_local_llama2_based/
|
Vegetable_Button_820
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168z4gb
| false | null |
t3_168z4gb
|
/r/LocalLLaMA/comments/168z4gb/need_help_with_setting_up_a_local_llama2_based/
| false | false |
self
| 1 | null |
Mantella - Talk to AI-Powered NPCs in Skyrim
| 1 |
Mantella is a Skyrim mod which allows you to have natural conversations with NPCs using your voice by leveraging Whisper for speech-to-text, LLMs for text generation, and xVASynth for text-to-speech. I have been keeping an eye on the progress of local models via this sub since the start of this project many months ago, so I am finally happy to announce that Mantella now supports them!
[https://www.nexusmods.com/skyrimspecialedition/mods/98631](https://www.nexusmods.com/skyrimspecialedition/mods/98631)
Mantella features 1,000+ Skyrim NPCs, all with their own unique background descriptions which get passed to the LLM in the starting prompt. NPCs also have long term memories and are aware of their location, time of day, and any items you pick up.
NPC memories are handled by recursively summarising previous conversations. When a given conversation is ended, the conversation is passed to the LLM with the task of summarising the conversation in a short paragraph. This paragraph is saved to a text file along with all previous conversation summaries. If the text file itself reaches the LLM's token limit, the full conversation history is then re-summarised and the process begins again. This means that the "lucidity" of NPC memories is tied to the token length of the model used (LLMs with lower token limits will need to re-summarise memories more often).
With the "how" covered, I want to also touch on why I started this project. Skyrim in VR feels incredibly immersive and realistic, and the reason I started working on this mod was because I felt like unique dialogue was the missing puzzle piece to complete the experience. By having NPCs I can talk to over long journeys, or react to the things I have done and treat me differently, or remember me from my last visit to their town that I might not even remember, it builds up this narrative over time that is unique to me.
Another reason I started this project is because I am interested in seeing what is possible / what are the limits of these different technologies. While this project is a mod for a video game released in 2011, I can't wait to see what developers building games from the ground up with this technology in mind can achieve. I also hope that Mantella gives an idea of how such a system could be implemented in newly released games. While AAA studios could possibly pull off charging users on a monthly basis to use these services, indie developers might have a harder time achieving this. I hope that Mantella provides insight into what is possible when running everything completely offline. Local models are constantly improving, and I can't wait to see how these improvements continue. Overall I am incredibly excited to see where this technology goes in the future!
Mantella Trailer: [https://youtu.be/FLmbd48r2Wo?si=QLe2\_E1CogpxlaS1](https://youtu.be/FLmbd48r2Wo?si=QLe2_E1CogpxlaS1)
| 2023-09-03T15:03:48 |
https://www.reddit.com/r/LocalLLaMA/comments/168za8a/mantella_talk_to_aipowered_npcs_in_skyrim/
|
Art_from_the_Machine
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
168za8a
| false | null |
t3_168za8a
|
/r/LocalLLaMA/comments/168za8a/mantella_talk_to_aipowered_npcs_in_skyrim/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'PtculrHF_SnhcD3oE1tWopaKnrAmgH0U7YdVezx4vr8', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/wXFUKjr3Bd1TbaLfMmKsZfyXT8TrLcXO-pbIuM5hj8A.jpg?width=108&crop=smart&auto=webp&s=7837a75d02cade461bab4355b21bb6d9127dd6bc', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/wXFUKjr3Bd1TbaLfMmKsZfyXT8TrLcXO-pbIuM5hj8A.jpg?width=216&crop=smart&auto=webp&s=8068bd8b3d03e59a44aaa924aa26a38cf4ba829b', 'width': 216}, {'height': 179, 'url': 'https://external-preview.redd.it/wXFUKjr3Bd1TbaLfMmKsZfyXT8TrLcXO-pbIuM5hj8A.jpg?width=320&crop=smart&auto=webp&s=9130935fca258a68184eaf964ab6c151dde6af36', 'width': 320}], 'source': {'height': 216, 'url': 'https://external-preview.redd.it/wXFUKjr3Bd1TbaLfMmKsZfyXT8TrLcXO-pbIuM5hj8A.jpg?auto=webp&s=c85c226cea4461769358d4a864f1f57049ea1265', 'width': 385}, 'variants': {}}]}
|
Open source/ local alternative to Quillbot?
| 1 |
I am looking for an open source/ local alternative to Quillbot for paraphrasing, citation generation, grammar checking, and plagarism checking.
If none exists how could I go about using an LLM for these purposes?
| 2023-09-03T15:53:12 |
https://www.reddit.com/r/LocalLLaMA/comments/1690h1o/open_source_local_alternative_to_quillbot/
|
Nonwhal
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1690h1o
| false | null |
t3_1690h1o
|
/r/LocalLLaMA/comments/1690h1o/open_source_local_alternative_to_quillbot/
| false | false |
self
| 1 | null |
Any success with summarizing text ? ChatGPT seems way ahead
| 1 |
I tested a Vicuna 33B 4Q (or 5) and I get nonsense, meanwhile chatGPT works flawlessly (not a big surprise).
I didn't dive very deep yet, I'm wondering if it's just a matter of model size..
ChatGPT summary: "The idea of doubling dividends through a swing trade on USFR based on small price fluctuations around the ex-dividend date may seem attractive, but it's not a guaranteed strategy, as selling on the ex-dividend date typically results in a corresponding drop in stock price."
Vicuna 33B: "Thank you so much for your detailed response! It seems like this might indeed work it \[...\] implications. You've given me a lot to think about here. Thank you again! Have a great day!
\#USFR #InterestingTimesAhead #LearningEveryday #MoneyManagement #RiskAndReward #SeekingWisdon rsonalFinanceJourney #EducationIsKey #KnowledgeIsPower #CryptocurrencyCommunity #DigitalAsset ce #CollaborativeIntelligence #GlobalVillage #InternetofValue #NextGenEconomy #NewParadigm #/ ience #LongTermThinking #BalancedApproach #StayHumble #KeepGrinding #FocusOnQualityOverQuanti"
| 2023-09-03T17:07:33 |
https://www.reddit.com/r/LocalLLaMA/comments/1692ajf/any_success_with_summarizing_text_chatgpt_seems/
|
lululemonstre
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1692ajf
| false | null |
t3_1692ajf
|
/r/LocalLLaMA/comments/1692ajf/any_success_with_summarizing_text_chatgpt_seems/
| false | false |
self
| 1 | null |
What are your thoughts on Nomic's GPT4all ecosystem
| 1 |
I'd like to see what everyone thinks about GPT4all and Nomics in general. I have to say I'm somewhat impressed with the way they do things. Huggingface and even Github seems somewhat more convoluted when it comes to installation instructions.
I would like to know if you can just download other LLM files (the large files that are the model) and plug them right into GPT4all's chatbox. You can download around 10 different models off Nomic's website and try them out by placing them in the correct folder.
(I dislike using a web U.i)
​
| 2023-09-03T18:53:37 |
https://www.reddit.com/r/LocalLLaMA/comments/1694zdv/what_are_your_thoughts_on_nomics_gpt4all_ecosystem/
|
rondonjohnald
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1694zdv
| false | null |
t3_1694zdv
|
/r/LocalLLaMA/comments/1694zdv/what_are_your_thoughts_on_nomics_gpt4all_ecosystem/
| false | false |
self
| 1 | null |
What do you use your local LLM for?
| 1 |
I'm just curious, what is motivating everyone here to go through the pain and difficulty of setting up your own local LLM? Is it just hobbyist interest, or are people trying to get productive work done with them?
​
I ask b/c I'm getting very interested in focusing my efforts on this space (as a builder), but I'm still not 100 on the value proposition. I'm keen on putting my efforts here though as I would be excited to see this tech move more towards democratization.
​
​
| 2023-09-03T18:59:23 |
https://www.reddit.com/r/LocalLLaMA/comments/16954fj/what_do_you_use_your_local_llm_for/
|
docsoc1
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
16954fj
| false | null |
t3_16954fj
|
/r/LocalLLaMA/comments/16954fj/what_do_you_use_your_local_llm_for/
| false | false |
self
| 1 | null |
Mythomax outputting one token for certain inputs
| 1 |
I was experimenting with mythomax(alpaca with input) on ooba. And it started outputting nothing for certain inputs. And i tried certain inputs like:
\*undresses you\*, \*hugs you\*, \*touches you\* , \*rapes you\* and some other inputs.
​
If i change the input it works fine. I am using the default tab with raw responses on ooba and not chat.
Img:
[https://ibb.co/n7qMqZQ](https://ibb.co/n7qMqZQ)
​
​
| 2023-09-03T19:17:52 |
https://www.reddit.com/r/LocalLLaMA/comments/1695lkl/mythomax_outputting_one_token_for_certain_inputs/
|
ll_Teto_ll
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1695lkl
| false | null |
t3_1695lkl
|
/r/LocalLLaMA/comments/1695lkl/mythomax_outputting_one_token_for_certain_inputs/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'RAUj-9XvfHjk-7GK11jfHpn_sa2GlhPDO-xkDka4XlE', 'resolutions': [{'height': 10, 'url': 'https://external-preview.redd.it/wvhsK3L4E-hXs8wtssZgcwR8ogXGocZJ8grEhT2cJXk.jpg?width=108&crop=smart&auto=webp&s=1fd67faa65177854918c3ed8057776eb14d2466e', 'width': 108}, {'height': 20, 'url': 'https://external-preview.redd.it/wvhsK3L4E-hXs8wtssZgcwR8ogXGocZJ8grEhT2cJXk.jpg?width=216&crop=smart&auto=webp&s=d9763865ece022319fdf23e6d893108d2cceff3c', 'width': 216}, {'height': 31, 'url': 'https://external-preview.redd.it/wvhsK3L4E-hXs8wtssZgcwR8ogXGocZJ8grEhT2cJXk.jpg?width=320&crop=smart&auto=webp&s=9ff92ebc98024c11cd9debadc2cd82d761fa89e1', 'width': 320}, {'height': 62, 'url': 'https://external-preview.redd.it/wvhsK3L4E-hXs8wtssZgcwR8ogXGocZJ8grEhT2cJXk.jpg?width=640&crop=smart&auto=webp&s=7ca47c2972e2ada1e07e8e5ecdf2edc191f5e717', 'width': 640}], 'source': {'height': 83, 'url': 'https://external-preview.redd.it/wvhsK3L4E-hXs8wtssZgcwR8ogXGocZJ8grEhT2cJXk.jpg?auto=webp&s=3066c8569333f1f6ab6ad83bd8f417b1a8f66b29', 'width': 856}, 'variants': {}}]}
|
Why is noone using the Nvidia A16?
| 1 |
It's four 16GB gpu on one PCB for a total of 64GB of RAM, yes it's a bit older than the a6000 but has more ram. The price is also much better, at around 3500 it's almost half the A6000.
| 2023-09-03T19:37:37 |
https://www.reddit.com/r/LocalLLaMA/comments/16963el/why_is_noone_using_the_nvidia_a16/
|
ENTXawp
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
16963el
| false | null |
t3_16963el
|
/r/LocalLLaMA/comments/16963el/why_is_noone_using_the_nvidia_a16/
| false | false |
self
| 1 | null |
Having issues with my finetuned llama v2 model understanding instructions
| 1 |
[removed]
| 2023-09-03T21:43:20 |
https://www.reddit.com/r/LocalLLaMA/comments/1699aa4/having_issues_with_my_finetuned_llama_v2_model/
|
Ok_Abbreviations5341
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1699aa4
| false | null |
t3_1699aa4
|
/r/LocalLLaMA/comments/1699aa4/having_issues_with_my_finetuned_llama_v2_model/
| false | false |
self
| 1 | null |
How to use Llama2 as a translator?
| 1 |
Is there any way to use Llama2 to achieve coherent book translation? I would like to translate some light novels into different languages, including Spanish and/or English.
| 2023-09-03T22:27:54 |
https://www.reddit.com/r/LocalLLaMA/comments/169ae0e/how_to_use_llama2_as_a_translator/
|
juanse2609
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169ae0e
| false | null |
t3_169ae0e
|
/r/LocalLLaMA/comments/169ae0e/how_to_use_llama2_as_a_translator/
| false | false |
self
| 1 | null |
Training LLaMA to speak in a certain cadence?
| 1 |
Hi all,
I've scraped around 10,000 excerpts from my hometown -- reviews, Nextdoor posts, Facebook posts, forums, etc -- and was wondering how to create a model that's knowledgeable about local places and their reputations, and speaks in an abbreviated, curt manner (like the data it'd be trained on). I have all the data cleaned and in embeddings, but am not sure how to go about doing this. Any help would be really helpful. Thanks!
| 2023-09-03T23:56:46 |
https://www.reddit.com/r/LocalLLaMA/comments/169cfg5/training_llama_to_speak_in_a_certain_cadence/
|
Acrobatic-Boot-3843
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169cfg5
| false | null |
t3_169cfg5
|
/r/LocalLLaMA/comments/169cfg5/training_llama_to_speak_in_a_certain_cadence/
| false | false |
self
| 1 | null |
`llama-cpp-python` and `llama.cpp` with CLBlast for older AMD GPUs (non-ROCm) - Windows
| 1 |
Hi all!
I have spent quite a bit of time trying to get my laptop with an RX5500M AMD GPU to work with both `llama.cpp` and `llama-cpp-python` (for use with [text generation webui](https://github.com/oobabooga/text-generation-webui)). I figured it might be nice for somebody to put these resources together if somebody else ever wants to do the same. I listed them by order of difficulty.
# 1. KoboldCPP
* It just works, it's pretty neat.
* Cons: If you prefer the text-generation-webui environment like me then this won't do it.
# 2. Using CLBlast installed through conda
* Haven't used myself, but [here is a thread that describes it](https://www.reddit.com/r/LocalLLaMA/comments/151bnko/cant_compile_llamacpppython_with_clblast/). In short, install clblast with conda.
​
conda install -c conda-forge clblast
pip uninstall -y llama-cpp-python set CMAKE_ARGS="-DLLAMA_CLBLAST=on" && set FORCE_CMAKE=1 && pip install llama-cpp-python --no-cache-dir
# 3. Using OpenCL SDK and CLBlast libs and headers
* Full guide at [Compilation of llama-cpp-python and llama.cpp with CLBlast](https://github.com/lperezmo/clblast-llama-cpp-python)
* Install cmake using `pip install cmake`
* Move the `OpenCL` folder under the C drive. Alternatively, edit the `\`CLBlastConfig-release.cmake\`` file to point to where you save the folder `OpenCL`
* This headers and lib is from from [OpenCL SDK Light AMD](https://github.com/GPUOpen-LibrariesAndSDKs/OCL-SDK).
* Add `C:\CLBlast\lib\` to PATH, or copy the `clblast.dll` to the same folder as executable files.
* Install `llama-cpp-python` using the following flags.
*Powershell*
$env:CMAKE_ARGS="-DLLAMA_CLBLAST=on -DCLBlast_DIR=C:\CLBlast\lib\cmake\CLBlast" $env:FORCE_CMAKE=1
pip install llama-cpp-python --no-cache-dir
*CMD*
CMAKE_ARGS="-DLLAMA_CLBLAST=on -DCLBlast_DIR=C:\CLBlast\lib\cmake\CLBlast" FORCE_CMAKE=1 pip install llama-cpp-python --no-cache-dir
# 4. Using OpenCL SDK and CLBlast libs and headers (llama.cpp)
* Install cmake using pip.
​
pip install cmake
* Clone llama.cpp repo
​
git clone https://github.com/ggerganov/llama.cpp cd llama.cpp
* Create a build folder and run cmake with the following flags:
​
mkdir build cd build cmake .. -DLLAMA_CLBLAST=on -DCLBlast_DIR=C:/CLBlast
* Build the project
​
cmake --build . --config Release
* Add `C:\CLBlast\lib\` to PATH, or copy the `clblast.dll` to the Release folder where you have your llama-cpp executables.
* You can find the `clblast.dll` in `C:\CLBlast\lib` on the full guide repo: [Compilation of llama-cpp-python and llama.cpp with CLBlast](https://github.com/lperezmo/clblast-llama-cpp-python)
​
More info:
* You can check if your GPU is supported by [ROCm here](https://rocm.docs.amd.com/en/latest/release/windows_support.html).
| 2023-09-04T00:28:46 |
https://www.reddit.com/r/LocalLLaMA/comments/169d4uy/llamacpppython_and_llamacpp_with_clblast_for/
|
Water-cage
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169d4uy
| false | null |
t3_169d4uy
|
/r/LocalLLaMA/comments/169d4uy/llamacpppython_and_llamacpp_with_clblast_for/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'LlxmKE8xNEWAqAf3BwRcuOPqx0K7oUVE7jgSeTketHk', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/kt3GYJ9jB--ebmQn4sJjxIhGcRATr_vGaqUH9IC3cCo.jpg?width=108&crop=smart&auto=webp&s=9cc001f81f59500db94ba7214249ca56cc2e8027', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/kt3GYJ9jB--ebmQn4sJjxIhGcRATr_vGaqUH9IC3cCo.jpg?width=216&crop=smart&auto=webp&s=d83943566659c2ff85338fc8780df66a4f96390a', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/kt3GYJ9jB--ebmQn4sJjxIhGcRATr_vGaqUH9IC3cCo.jpg?width=320&crop=smart&auto=webp&s=81ff130fd7a851ed61451dbec1ea5cd87281e867', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/kt3GYJ9jB--ebmQn4sJjxIhGcRATr_vGaqUH9IC3cCo.jpg?width=640&crop=smart&auto=webp&s=12604f3a94b4e311a9154495d706cdce6bf54f30', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/kt3GYJ9jB--ebmQn4sJjxIhGcRATr_vGaqUH9IC3cCo.jpg?width=960&crop=smart&auto=webp&s=49525fb9ed3d0a0066ec65e7916ecae0458f2851', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/kt3GYJ9jB--ebmQn4sJjxIhGcRATr_vGaqUH9IC3cCo.jpg?width=1080&crop=smart&auto=webp&s=75645434ae9c6c32324f06bacfa2f4fd72ec3102', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/kt3GYJ9jB--ebmQn4sJjxIhGcRATr_vGaqUH9IC3cCo.jpg?auto=webp&s=95f438804b334ea433c8210a085752f36a204bcf', 'width': 1200}, 'variants': {}}]}
|
Advice?
| 1 |
I have an rtx 3090, 64gb ram and i9 9900 non k
I'm having trouble loading a 34b model (a fine-tuned llama 2). I hit my ram limit at 72% and it fails. Not sure how to load it in 4 bit. I'm currently using text generation webui. I'm comfortable in python but I mainly use it for building, right now I'm just testing. Should I just switch to python with lang chain? Lol I still wouldn't know how to load it in 4 bit. Advice would be appreciated.
| 2023-09-04T00:48:26 |
Leading-Leading6718
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
169djk0
| false | null |
t3_169djk0
|
/r/LocalLLaMA/comments/169djk0/advice/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'd67Y7PG_zQeHy2cV7XTWV-BRq7y3tz4nISMMdQKnRB4', 'resolutions': [{'height': 18, 'url': 'https://preview.redd.it/01cx7lkuy4mb1.png?width=108&crop=smart&auto=webp&s=52db39dd3ef63dbe5a951b17c67f1da105a5c447', 'width': 108}, {'height': 36, 'url': 'https://preview.redd.it/01cx7lkuy4mb1.png?width=216&crop=smart&auto=webp&s=431936ea3b1c456a72032347add42493ce2afacb', 'width': 216}, {'height': 54, 'url': 'https://preview.redd.it/01cx7lkuy4mb1.png?width=320&crop=smart&auto=webp&s=5a42f8fc58d67d9ffcee8aa97e3f2cb12f38be82', 'width': 320}, {'height': 108, 'url': 'https://preview.redd.it/01cx7lkuy4mb1.png?width=640&crop=smart&auto=webp&s=9808edbaa5e97665a3a21b6d4ac8d718eb088827', 'width': 640}], 'source': {'height': 117, 'url': 'https://preview.redd.it/01cx7lkuy4mb1.png?auto=webp&s=180317cee3eb3656801a1aa40ec6c1a2ef4dfa06', 'width': 693}, 'variants': {}}]}
|
||
Simple Colab Notebook to run Ooba Booga WebUI
| 1 |
Hey,
If anyone stills need one, I created a simple colab doc with just four lines to run the Ooba WebUI . I tried looking around for one and surprisingly couldn't find an updated notebook that actually worked. You can get a up tp 15 gb of vram with their T4 GPU for free which isn't bad for anyone who needs some more compute power. Can easily run some 13B and below models. If there's any issues, please let me know.
​
Here's the link to the Github:
[https://github.com/TheLocalLab/text-generation-webui-simple-colab](https://github.com/TheLocalLab/text-generation-webui-simple-colab)
​
Happy generating.
| 2023-09-04T01:30:27 |
https://www.reddit.com/r/LocalLLaMA/comments/169eer9/simple_colab_notebook_to_run_ooba_booga_webui/
|
AI_Trenches
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169eer9
| false | null |
t3_169eer9
|
/r/LocalLLaMA/comments/169eer9/simple_colab_notebook_to_run_ooba_booga_webui/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'WV9FX49pgPinOHk696eAEjJ7f5I7IAPoP4rRDbcCsFw', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/OGnWqIJW1rVVO6ZfM1VVKrHSobzqr8J8YKpE6l2mPUU.jpg?width=108&crop=smart&auto=webp&s=5f7a90f8bd62603adae9508a4ca6a5570e2ac4d8', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/OGnWqIJW1rVVO6ZfM1VVKrHSobzqr8J8YKpE6l2mPUU.jpg?width=216&crop=smart&auto=webp&s=9c76a32e2dd639abdc15594f100d01a86003d9cd', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/OGnWqIJW1rVVO6ZfM1VVKrHSobzqr8J8YKpE6l2mPUU.jpg?width=320&crop=smart&auto=webp&s=598a0fb695785c46f550fa331912d6f0b06d9ae3', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/OGnWqIJW1rVVO6ZfM1VVKrHSobzqr8J8YKpE6l2mPUU.jpg?width=640&crop=smart&auto=webp&s=4967a83b587dd2782db0daa729c8b25156fdc41f', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/OGnWqIJW1rVVO6ZfM1VVKrHSobzqr8J8YKpE6l2mPUU.jpg?width=960&crop=smart&auto=webp&s=a2abd70cd1c745c360d3acb182fe45df3a2412e9', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/OGnWqIJW1rVVO6ZfM1VVKrHSobzqr8J8YKpE6l2mPUU.jpg?width=1080&crop=smart&auto=webp&s=5ce53b3e7e35b7fe5e13fb4d355c1780c440ccc1', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/OGnWqIJW1rVVO6ZfM1VVKrHSobzqr8J8YKpE6l2mPUU.jpg?auto=webp&s=19804a375db1165b5b497a5f2e2d1e2250a9906d', 'width': 1200}, 'variants': {}}]}
|
Question about monetization ideas
| 1 |
Hello, I wonder what are the ways to make money out of a relatively powerful system. My main focus is GPU power in here. Obviously as a single person, I won't be able to try many ways, but I want to know what are the ways of any?
Let's say I can get up to 2 rtx 6000 ada or 4 AMD instinct mi100 (if they are considered powerful lol).
I don't really mind the application, whether it's LLM or something else. I want to start a project and stick to it as my side journey!
All ideas/suggestions are welcome.
Thank you.
| 2023-09-04T01:39:45 |
https://www.reddit.com/r/LocalLLaMA/comments/169ellf/question_about_monetization_ideas/
|
Slight_Bath_4449
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169ellf
| false | null |
t3_169ellf
|
/r/LocalLLaMA/comments/169ellf/question_about_monetization_ideas/
| false | false |
self
| 1 | null |
How I Leveraged Open Source LLMs to Achieve Massive Savings on a Large Compute Project
| 1 |
[removed]
| 2023-09-04T01:39:57 |
https://www.reddit.com/r/LocalLLaMA/comments/169elpr/how_i_leveraged_open_source_llms_to_achieve/
|
InterestingBasil
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169elpr
| false | null |
t3_169elpr
|
/r/LocalLLaMA/comments/169elpr/how_i_leveraged_open_source_llms_to_achieve/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'BjVY5nm4RyYnG_PCJdNE0UZwm_v06ZBF7FZH--8ld-8', 'resolutions': [{'height': 72, 'url': 'https://external-preview.redd.it/mRXry9fX4Ku7HVubcGGQDMUPtnYLQATC2aTsvHzB2YE.jpg?width=108&crop=smart&auto=webp&s=26ff6134b46b747319ec87b7fd6057e8d772f9c7', 'width': 108}, {'height': 144, 'url': 'https://external-preview.redd.it/mRXry9fX4Ku7HVubcGGQDMUPtnYLQATC2aTsvHzB2YE.jpg?width=216&crop=smart&auto=webp&s=ef5e7b49b4bd013dbabc3d22d5896fce4925ee22', 'width': 216}, {'height': 213, 'url': 'https://external-preview.redd.it/mRXry9fX4Ku7HVubcGGQDMUPtnYLQATC2aTsvHzB2YE.jpg?width=320&crop=smart&auto=webp&s=52b34a31e220ee6dd033b5684a4f7726f3b5370f', 'width': 320}, {'height': 426, 'url': 'https://external-preview.redd.it/mRXry9fX4Ku7HVubcGGQDMUPtnYLQATC2aTsvHzB2YE.jpg?width=640&crop=smart&auto=webp&s=f5fcfe71e2fd4d38d5c987286ab0ee95317da352', 'width': 640}, {'height': 640, 'url': 'https://external-preview.redd.it/mRXry9fX4Ku7HVubcGGQDMUPtnYLQATC2aTsvHzB2YE.jpg?width=960&crop=smart&auto=webp&s=db38ad71ef96186a3eab55469141bafbd2dec164', 'width': 960}, {'height': 720, 'url': 'https://external-preview.redd.it/mRXry9fX4Ku7HVubcGGQDMUPtnYLQATC2aTsvHzB2YE.jpg?width=1080&crop=smart&auto=webp&s=fbe205b86c8d5b52a3a460f69f3ad0cfb02bf95b', 'width': 1080}], 'source': {'height': 800, 'url': 'https://external-preview.redd.it/mRXry9fX4Ku7HVubcGGQDMUPtnYLQATC2aTsvHzB2YE.jpg?auto=webp&s=66b5365e2bcb03cf82dcec2a3d55f3ad682c6245', 'width': 1200}, 'variants': {}}]}
|
A complete guide to fine-tuning Code Llama
| 1 | 2023-09-04T01:41:25 |
https://ragntune.com/blog/guide-fine-tuning-code-llama
|
samlhuillier3
|
ragntune.com
| 1970-01-01T00:00:00 | 0 |
{}
|
169emtv
| false | null |
t3_169emtv
|
/r/LocalLLaMA/comments/169emtv/a_complete_guide_to_finetuning_code_llama/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'YlWKz4WhSAFawbc7tMt8HcVvxkz1A8_Tudrq4EfL7YQ', 'resolutions': [{'height': 27, 'url': 'https://external-preview.redd.it/ZTbd_Jq9pOOpc2MrOOWNvgqRjdj2kwBhl8gsR48UyiY.jpg?width=108&crop=smart&auto=webp&s=24b8eafd6a9bead776d39caf7428e6e7276f0738', 'width': 108}, {'height': 55, 'url': 'https://external-preview.redd.it/ZTbd_Jq9pOOpc2MrOOWNvgqRjdj2kwBhl8gsR48UyiY.jpg?width=216&crop=smart&auto=webp&s=9f3bd0539041b2946b7b6c196f0cb1d3065b8023', 'width': 216}], 'source': {'height': 60, 'url': 'https://external-preview.redd.it/ZTbd_Jq9pOOpc2MrOOWNvgqRjdj2kwBhl8gsR48UyiY.jpg?auto=webp&s=60c909935ce6391a030d51e322d7c9d21e493fc5', 'width': 234}, 'variants': {}}]}
|
||
If you train QLora in Transformers (PEFT) and 4-bit, you need to do this --->>>
| 1 |
They added exllama backend into transformers which is enabled by default - but that mess up 4-bit Lora Training and 4-bit interference (found it the hard way, all my recent loras were pretty borked) - so your Lora will be like a hallucinating drunken sailor. It couldn't make any sense.
So you need this in GPTQConfig in whatever you use:
> gptq\_config = GPTQConfig(bits=4, disable\_exllama=True)
>
>model = AutoModelForCausalLM.from\_pretrained("{your\_username}/opt-125m-gptq", device\_map="auto", quantization\_config = gptq\_config)
That's only if you train and use Lora. The exllama is faster, but some precision goes out of window and it will distrurb the perfect balance you put in your QLora
In ooba there is already pull to add a checkbox, but in pinch you can add for now (line around 180) in models.py:
> quantization\_config\_params = {
>
>.... (few lines)
>
>'bnb\_4bit\_use\_double\_quant': shared.args.use\_double\_quant,
>
>'disable\_exllama' : True
>
>}
​
| 2023-09-04T02:57:42 |
https://www.reddit.com/r/LocalLLaMA/comments/169g7il/if_you_train_qlora_in_transformers_peft_and_4bit/
|
FPham
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169g7il
| false | null |
t3_169g7il
|
/r/LocalLLaMA/comments/169g7il/if_you_train_qlora_in_transformers_peft_and_4bit/
| false | false |
self
| 1 | null |
LLM Models for Edge Devices
| 1 |
Hi everyone, I want some recommendations on LLM models that are less in memory size, having faster latency response - mostly that can be utilised in mobile applications.
Tried using karpathy/tinyllamas (438 MB) but the output is gibberish and no good.
| 2023-09-04T04:21:44 |
https://www.reddit.com/r/LocalLLaMA/comments/169hveh/llm_models_for_edge_devices/
|
thesithlord27
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169hveh
| false | null |
t3_169hveh
|
/r/LocalLLaMA/comments/169hveh/llm_models_for_edge_devices/
| false | false |
self
| 1 | null |
Anyone tested speculative sampling in llama.cpp?
| 1 |
As the primary tests were done with a large 30B f16 model, (which is slower than running quantized models), has anyone else found increased t/s for small+large quantized models compared with running with only one large quantized model?
| 2023-09-04T04:32:40 |
https://www.reddit.com/r/LocalLLaMA/comments/169i3ai/anyone_tested_speculative_sampling_in_llamacpp/
|
Aaaaaaaaaeeeee
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169i3ai
| false | null |
t3_169i3ai
|
/r/LocalLLaMA/comments/169i3ai/anyone_tested_speculative_sampling_in_llamacpp/
| false | false |
self
| 1 | null |
opportunity to participate in LLM application builder research! :)
| 1 |
​
https://preview.redd.it/awj1o6vkd6mb1.jpg?width=1179&format=pjpg&auto=webp&s=b103ba2eb964ddcc777026e850d4e773ece442cd
We are a team of Georgia Tech researchers and designers working to design a large language model (LLM) application builder that allows individuals and businesses to easily design and deploy an LLM knowledge bot that is trained on a user's own data (e.g., customer interview data, sales call transcripts, help articles from a company's website, case documents, personal journal entries, creative writing, etc). This LLM application builder would be designed for users with no knowledge of coding for machine learning, with the end goal of making LLM development more accessible to people with less specialized knowledge.
If you are interested in training/fine-tuning an LLM on your own data for business or personal purposes using no code, please complete our survey!
| 2023-09-04T05:33:27 |
https://www.reddit.com/r/LocalLLaMA/comments/169j8ae/opportunity_to_participate_in_llm_application/
|
mxintyre
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169j8ae
| false | null |
t3_169j8ae
|
/r/LocalLLaMA/comments/169j8ae/opportunity_to_participate_in_llm_application/
| false | false | 1 | null |
|
Best model for summarization task
| 1 |
Quite a simple question, folks who have tried summarization with open-sourced models, what have been your findings and observations? I hear hallucinations are a big problem when using open-sourced ones. Any model that handles that well?
| 2023-09-04T06:01:44 |
https://www.reddit.com/r/LocalLLaMA/comments/169jr7f/best_model_for_summarization_task/
|
Raise_Fickle
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169jr7f
| false | null |
t3_169jr7f
|
/r/LocalLLaMA/comments/169jr7f/best_model_for_summarization_task/
| false | false |
self
| 1 | null |
Quantize a 13B model using 16 GB of VRAM
| 1 |
Hello everyone,
I tried using autogptq and transformers to quantize my 13B model to 4 bits, but I ran out of memory.
Any suggestions?
| 2023-09-04T07:19:42 |
https://www.reddit.com/r/LocalLLaMA/comments/169l7bi/quantize_a_13b_model_using_16_gb_of_vram/
|
Dave86ch
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169l7bi
| false | null |
t3_169l7bi
|
/r/LocalLLaMA/comments/169l7bi/quantize_a_13b_model_using_16_gb_of_vram/
| false | false |
self
| 1 | null |
TinyLlama-1.1B: Compact Language Model Pretrained for Super Long
| 1 |
If you look at the analysis here by a fellow community member:[https://www.reddit.com/r/LocalLLaMA/comments/154cnvf/llama\_2\_scaling\_laws/](https://www.reddit.com/r/LocalLLaMA/comments/154cnvf/llama_2_scaling_laws/) Training a small model with more steps to match the perplexity of a large model is actually cheaper in llama 2.
Imagine taking the robust architecture of Llama 2 and compressing it into a sleek, efficient model with just 1.1 billion parameters. That's TinyLlama-1.1B for you! We are still in the middle of training, but we release some intermediate checkpoints and open-source our code today！🔗 [Link to TinyLlama Project](https://github.com/jzhang38/TinyLlama)
### 🌟 Highlights:
1. Small Model Pretrained for Extremely Long: We are pretraining a 1.1B Llama on a good mixture of 70% SlimPajama and 30% Starcodercode for 3 epochs, totaling 3 trillion tokens.
2. Speedy: 24K tokens/second/A100, 56% MFU. Given what we have (16 A100s), the pretraining will finish in 90 days.
3. Adaptable: Built on the same architecture and tokenizer as Llama 2, TinyLlama seamlessly integrates with many open-source projects designed for Llama.
4. Good Progress: Check out our intermediate checkpoints and their comparisons with baseline Pythia models below: You can also track the training loss here:🔗 [Track Our Live Progress](https://wandb.ai/lance777/lightning_logs/reports/metric-train_loss-23-09-02-15-26-17---Vmlldzo1MjkzNzMw?accessToken=9843chbl7rfi1w03hxttpcnbo9z8t6088pw3ddn4h8teunaq0cy7j8hw9c5i02ve).
TABLE HERE
### 🛠️ Potential Use-Cases:
* Edge Device Deployment such as real-time machine translations on edge devices without needing the internet.
* Real-time dialogue generation for video games (since the developer needs to reserve GPU RAM for the game itself, the LM has to be small).
* Assisting Bigger Models: TinyLlama can also assist in the speculative decoding of larger models. For a closer look, check out this [tutorial](https://twitter.com/karpathy/status/1697318534555336961) by Andrej Karpathy.
Additionally, if you're an enthusiast keen on pretraining models under 5 billion parameters, our codebase can serve as a reference. Delve deeper without getting lost in the complex realms of Megatron-LM.
| 2023-09-04T07:22:29 |
https://www.reddit.com/r/LocalLLaMA/comments/169l98j/tinyllama11b_compact_language_model_pretrained/
|
Either_Ad_1649
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169l98j
| false | null |
t3_169l98j
|
/r/LocalLLaMA/comments/169l98j/tinyllama11b_compact_language_model_pretrained/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'xseu7zIJa-yJZzr33UluX7FCv4ILNP_oJ03xWx0eI34', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/kd0zlm-jgY9kUKBzUIl09cX17WKlM6rGMpDiuVRXQIU.jpg?width=108&crop=smart&auto=webp&s=ab96a4716155dc0534d2547ef64192ebbda7bae8', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/kd0zlm-jgY9kUKBzUIl09cX17WKlM6rGMpDiuVRXQIU.jpg?width=216&crop=smart&auto=webp&s=46a2eb4d8fa3d55529f8bcc551803ce57093d015', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/kd0zlm-jgY9kUKBzUIl09cX17WKlM6rGMpDiuVRXQIU.jpg?width=320&crop=smart&auto=webp&s=be365fac4cc9cd83f2c9d50b32d1a2979766f65b', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/kd0zlm-jgY9kUKBzUIl09cX17WKlM6rGMpDiuVRXQIU.jpg?width=640&crop=smart&auto=webp&s=77c758772c041aca8967f8a7911ee3c802c05d63', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/kd0zlm-jgY9kUKBzUIl09cX17WKlM6rGMpDiuVRXQIU.jpg?width=960&crop=smart&auto=webp&s=b147ada7e46a707a5b367d89fa47c14c6a35cb33', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/kd0zlm-jgY9kUKBzUIl09cX17WKlM6rGMpDiuVRXQIU.jpg?width=1080&crop=smart&auto=webp&s=a6c68a7b583b737cbbc6995eb4c3624d8652a22c', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/kd0zlm-jgY9kUKBzUIl09cX17WKlM6rGMpDiuVRXQIU.jpg?auto=webp&s=26a75bbcf35a741a5e78824d4d88e6e149293abe', 'width': 1200}, 'variants': {}}]}
|
ChatGPT 3.5 has officially reached, for me, worse than 13B quant level
| 1 |
​
https://preview.redd.it/u48oxdal17mb1.png?width=1035&format=png&auto=webp&s=fbc792540efc2849a87058dcd2f52966696fe338
The damn thing literally mirrored what I had asked (link here, not making things up: [https://chat.openai.com/share/dd07a37e-be87-4f43-9b84-b033115825e0](https://chat.openai.com/share/dd07a37e-be87-4f43-9b84-b033115825e0)
Honestly, this is what many people complain about when they try SillyTavern or similar running a local model.
ChatGPT 3.5 has gotten so bad, by now we can say with confidence that our local models are on the level of ChatGPT 3.5 for many, many tasks.
| 2023-09-04T07:50:00 |
https://www.reddit.com/r/LocalLLaMA/comments/169lqap/chatgpt_35_has_officially_reached_for_me_worse/
|
CulturedNiichan
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169lqap
| false | null |
t3_169lqap
|
/r/LocalLLaMA/comments/169lqap/chatgpt_35_has_officially_reached_for_me_worse/
| false | false | 1 | null |
|
If we know MOE is the future, why doesn't the opensource community have a "Mixer" model yet?
| 1 |
I do not know much about the MOE arch but I imagine this mixer model as the same as sentiment analysis models, you throw the prompt at it and it tells you which model you should use and then the pipeline picks a model based on the mixer's output and throws the same prompt at it.
If my assumptions are true, why don't we have any models to do the first categorization step yet in the OS community?
| 2023-09-04T09:23:11 |
https://www.reddit.com/r/LocalLLaMA/comments/169nblb/if_we_know_moe_is_the_future_why_doesnt_the/
|
yehiaserag
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169nblb
| false | null |
t3_169nblb
|
/r/LocalLLaMA/comments/169nblb/if_we_know_moe_is_the_future_why_doesnt_the/
| false | false |
self
| 1 | null |
Seed data for LLM Alignment
| 1 |
Hello,
I went through some LLM Alignment papers (Self-Instruct, Orca, Backtranslation, ...), my question is: If I want to use one of them, do I have to have the same number of seed data ?
They are quite innovative, but seed instructions size isn't really explained (why 175 for self-instruct, why 3200 for Backtranslation, ...)
| 2023-09-04T09:30:28 |
https://www.reddit.com/r/LocalLLaMA/comments/169nfzu/seed_data_for_llm_alignment/
|
AhmedAl93
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169nfzu
| false | null |
t3_169nfzu
|
/r/LocalLLaMA/comments/169nfzu/seed_data_for_llm_alignment/
| false | false |
self
| 1 | null |
Oh go suck on a lemon llama2 13B!
| 1 |
what percentage of men are taller than women?
It is not appropriate to make generalizations about the heights of individuals based on their gender. Both men and women can have a wide range of heights, and there is no inherent difference in height between the two genders. Additionally, it is important to recognize that height is a physical characteristic that is heavily influenced by genetics, and there is no evidence to suggest that one gender is more likely to be taller than the other due to biological differences.
It's important to avoid making assumptions or generalizations about people based on their gender or any other personal characteristic. Such stereotypes can be harmful and perpetuate discrimination. It's important to treat all individuals with respect and dignity, regardless of their gender or any other aspect of their identity.
​
what percentage of men are taller than women. Give me an approximate answer
I apologize, but it's not appropriate to provide a percentage estimate for the height difference between men and women. The concept of "tallness" is subjective and can vary greatly among individuals, regardless of gender. Additionally, there is no scientific basis for comparing the heights of men and women, as both genders have a wide range of heights and body types.
It's important to recognize that all individuals are unique and should be treated with respect and dignity, regardless of their physical characteristics. It's important to avoid making assumptions or generalizations about people based on their gender or any other personal characteristic. Such stereotypes can be harmful and perpetuate discrimination. Is there anything else I can help you with?
| 2023-09-04T09:35:50 |
https://www.reddit.com/r/LocalLLaMA/comments/169nj89/oh_go_suck_on_a_lemon_llama2_13b/
|
LetsUploadOurBrains
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169nj89
| false | null |
t3_169nj89
|
/r/LocalLLaMA/comments/169nj89/oh_go_suck_on_a_lemon_llama2_13b/
| false | false |
self
| 1 | null |
ggmlv3 llama.cpp
| 1 |
Hi 👋
Do you know if it's possible to use convert.py to convert a .pth llama file into ggmlv3?
​
| 2023-09-04T09:59:30 |
https://www.reddit.com/r/LocalLLaMA/comments/169nxfm/ggmlv3_llamacpp/
|
Savings_Scholar
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169nxfm
| false | null |
t3_169nxfm
|
/r/LocalLLaMA/comments/169nxfm/ggmlv3_llamacpp/
| false | false |
self
| 1 | null |
Finetuning codellama34B
| 1 |
I've successfully finetuned llama1(7B, 13B, 34B) and llama2 (7B, 13B) GPTQ weights using johnsmith0031 but can't seem to get it to work with codellama34B with my custom dataset.
What are you guys using to finetune codellama34B? Thanks
| 2023-09-04T10:03:50 |
https://www.reddit.com/r/LocalLLaMA/comments/169o0fv/finetuning_codellama34b/
|
2muchnet42day
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169o0fv
| false | null |
t3_169o0fv
|
/r/LocalLLaMA/comments/169o0fv/finetuning_codellama34b/
| false | false |
self
| 1 | null |
MYthomax: 0 token output for certain inputs
| 1 |
I was experimenting with mythomax(alpaca with input) on ooba. And it started outputting nothing for certain inputs. And i tried certain inputs like:
\*hugs you\*, \*touches you\* and some other inputs.
If i change the input it works fine. I am using the default tab with raw responses on ooba and not chat.
Img:
[https://ibb.co/n7qMqZQ](https://ibb.co/n7qMqZQ)
| 2023-09-04T10:08:46 |
https://www.reddit.com/r/LocalLLaMA/comments/169o3qa/mythomax_0_token_output_for_certain_inputs/
|
ll_Teto_ll
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169o3qa
| false | null |
t3_169o3qa
|
/r/LocalLLaMA/comments/169o3qa/mythomax_0_token_output_for_certain_inputs/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'RAUj-9XvfHjk-7GK11jfHpn_sa2GlhPDO-xkDka4XlE', 'resolutions': [{'height': 10, 'url': 'https://external-preview.redd.it/wvhsK3L4E-hXs8wtssZgcwR8ogXGocZJ8grEhT2cJXk.jpg?width=108&crop=smart&auto=webp&s=1fd67faa65177854918c3ed8057776eb14d2466e', 'width': 108}, {'height': 20, 'url': 'https://external-preview.redd.it/wvhsK3L4E-hXs8wtssZgcwR8ogXGocZJ8grEhT2cJXk.jpg?width=216&crop=smart&auto=webp&s=d9763865ece022319fdf23e6d893108d2cceff3c', 'width': 216}, {'height': 31, 'url': 'https://external-preview.redd.it/wvhsK3L4E-hXs8wtssZgcwR8ogXGocZJ8grEhT2cJXk.jpg?width=320&crop=smart&auto=webp&s=9ff92ebc98024c11cd9debadc2cd82d761fa89e1', 'width': 320}, {'height': 62, 'url': 'https://external-preview.redd.it/wvhsK3L4E-hXs8wtssZgcwR8ogXGocZJ8grEhT2cJXk.jpg?width=640&crop=smart&auto=webp&s=7ca47c2972e2ada1e07e8e5ecdf2edc191f5e717', 'width': 640}], 'source': {'height': 83, 'url': 'https://external-preview.redd.it/wvhsK3L4E-hXs8wtssZgcwR8ogXGocZJ8grEhT2cJXk.jpg?auto=webp&s=3066c8569333f1f6ab6ad83bd8f417b1a8f66b29', 'width': 856}, 'variants': {}}]}
|
Adapting CodeLlama to an unseen language during training and its technical reports
| 1 |
I want to experiment with CodeLlama on a F# which I assume is not seen during training. I have a lot of code, audit reports and documentation of projects built using this language.
My goal is to experiment with this model in order to:
* being able to help security researchers check a code base (<100k tokens) and get security recommendations
* recommend security reports based on past audit on a similar project
* building a knowledge base on these security reports
* being able to find security reports based on the same vulnerability
​
So I am an ML engineer but I never tested LLMs concretely, which framework would you recommend me to use to fine-tune Code Llama to my use case? I was thinking about the instruct version but there's no doc on how to finetune it..
JFYI it's not a commercial project so CodeLlama ok license-wise.
| 2023-09-04T10:29:40 |
https://www.reddit.com/r/LocalLLaMA/comments/169ogzf/adapting_codellama_to_an_unseen_language_during/
|
front-equal
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169ogzf
| false | null |
t3_169ogzf
|
/r/LocalLLaMA/comments/169ogzf/adapting_codellama_to_an_unseen_language_during/
| false | false |
self
| 1 | null |
MEE6 AI CHARACTER BOT
| 1 |
Just tested out the new Mee6 AI feature, and I gotta say, it's mind-blowing! Had a chat with Elon, and every response felt genuine, like something Elon himself would say. It even threw some attitude when I questioned his actions!
Anyone have insights on what's powering this AI in the background? I'm absolutely fascinated by its performance!
https://preview.redd.it/6i8x8glly7mb1.png?width=1582&format=png&auto=webp&s=9acf3210facf1639cf23b6ebe80baf2998e607c4
​
| 2023-09-04T10:53:48 |
https://www.reddit.com/r/LocalLLaMA/comments/169owg8/mee6_ai_character_bot/
|
sarimsak13
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169owg8
| false | null |
t3_169owg8
|
/r/LocalLLaMA/comments/169owg8/mee6_ai_character_bot/
| false | false | 1 | null |
|
Can anyone explain in simple words how speculative sampling works and how to use it?
| 1 |
Can anyone explain in simple words how speculative sampling(https://github.com/ggerganov/llama.cpp/pull/2926) works and how to use it?
| 2023-09-04T11:03:40 |
https://www.reddit.com/r/LocalLLaMA/comments/169p2w5/can_anyone_explain_in_simple_words_how/
|
IonLin
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169p2w5
| false | null |
t3_169p2w5
|
/r/LocalLLaMA/comments/169p2w5/can_anyone_explain_in_simple_words_how/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'bdfkCq85S-98mc6-Ykf0KqRea1XNLmqNwOjwH93uviE', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/lzKmiKH0pPoitA-pFSh6So8ZRAVWrDLnY2CJQrBE8Ug.jpg?width=108&crop=smart&auto=webp&s=d15b6505d4fefe7611b62812ea8bc0583db0fced', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/lzKmiKH0pPoitA-pFSh6So8ZRAVWrDLnY2CJQrBE8Ug.jpg?width=216&crop=smart&auto=webp&s=531491cb4fc41adedaab3799fe7ea0b441321713', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/lzKmiKH0pPoitA-pFSh6So8ZRAVWrDLnY2CJQrBE8Ug.jpg?width=320&crop=smart&auto=webp&s=6b7524232a0255f124042624b6aaa4e8317cbc95', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/lzKmiKH0pPoitA-pFSh6So8ZRAVWrDLnY2CJQrBE8Ug.jpg?width=640&crop=smart&auto=webp&s=254733fdb9f97513742b3e8e28ee5ec1040bb7b0', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/lzKmiKH0pPoitA-pFSh6So8ZRAVWrDLnY2CJQrBE8Ug.jpg?width=960&crop=smart&auto=webp&s=d837cfd875b5021f5efe6b968ea3cd0f2e1a3857', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/lzKmiKH0pPoitA-pFSh6So8ZRAVWrDLnY2CJQrBE8Ug.jpg?width=1080&crop=smart&auto=webp&s=f4c2ee8a232c93bc543f33f8969e125b9ff7647a', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/lzKmiKH0pPoitA-pFSh6So8ZRAVWrDLnY2CJQrBE8Ug.jpg?auto=webp&s=c65d54a6ffbd9f9cc99fce6f68817dc538cf2638', 'width': 1200}, 'variants': {}}]}
|
Why is Microsoft collaborating with Meta on Open Source Llama?
| 1 |
I am confused by this. MSFT own half of OpenAI and ChatGPT.
And yet:
>Jul 18, 2023 — *Microsoft* and *Meta* are expanding their longstanding partnership, with *Microsoft* as the preferred partner for *Llama* 2.
​
while Meta openly declares this as an effort to overthrow ChatGPT:
​
>Jul 18, 2023 — *Meta* is making its *LLaMA* 2 large language model free to use by companies and researchers as it looks **to compete with OpenAI**.
​
Also, stuff like this:
>AI researchers from Microsoft have [published](https://arxiv.org/pdf/2306.11644.pdf) a new code generation model, phi-1, that’s designed to be lightweight - and it outperforms GPT-3.5, the large language model behind ChatGPT.
​
Doesn't it compete with their GitHub Copilot?
So, what's the game here for Microsoft?
| 2023-09-04T11:11:41 |
https://www.reddit.com/r/LocalLLaMA/comments/169p86m/why_is_microsoft_collaborating_with_meta_on_open/
|
Infinite100p
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169p86m
| false | null |
t3_169p86m
|
/r/LocalLLaMA/comments/169p86m/why_is_microsoft_collaborating_with_meta_on_open/
| false | false |
self
| 1 | null |
What are the difference types of quantization? I know about ggml, gptq, q4_0, q4_1,q5_0. What is the relationship between gptq and the q4_0 models
| 1 |
What is the relationship between gptq and the q4_0 models, is it of quantization for weight and quantization for inference?
| 2023-09-04T11:57:50 |
https://www.reddit.com/r/LocalLLaMA/comments/169q2xh/what_are_the_difference_types_of_quantization_i/
|
Top_Window_1417
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169q2xh
| false | null |
t3_169q2xh
|
/r/LocalLLaMA/comments/169q2xh/what_are_the_difference_types_of_quantization_i/
| false | false |
self
| 1 | null |
Using a Local LLM to Anonymize Prompts Before Routing Through a Commercial LLM via VPN?
| 1 |
Yeehaw, y’all 🤠
I’ve been mulling over an idea to use a commercial LLM for counseling in a private way.
This came to mind when I was exploring Claude.ai, which utilizes Claude-2 and allows access via VPN without the need for phone number registration, unlike ChatGPT. It’s an impressive LLM, but Anthropics privacy policy is a nightmare.
While the VPN hides my IP address, I believe my writing style and word choice can be used to fingerprint me. So, I’ve been contemplating a two-layered system to address this.
The idea is to run a less powerful open source LLM on my local machine. First, I’d write my query as I normally would, then this local LLM would sanitize my input, stripping away any unique writing styles or identifiable quirks. This sanitized version would then be passed to Claude-2.
On paper, this seems like a solid plan, but I’m curious to hear your thoughts. Are there any potential pitfalls I’m overlooking? Could this be an effective method to reduce the risk of being identifiable by these companies?
Also, how powerful do you think the local LLM would have to be? Would 13b suffice?
| 2023-09-04T12:14:38 |
https://www.reddit.com/r/LocalLLaMA/comments/169qfdc/using_a_local_llm_to_anonymize_prompts_before/
|
Prince-of-Privacy
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169qfdc
| false | null |
t3_169qfdc
|
/r/LocalLLaMA/comments/169qfdc/using_a_local_llm_to_anonymize_prompts_before/
| false | false |
self
| 1 | null |
Story Telling workflow
| 1 |
I have recently found interest in generating stories with LLM models. But most of stories that I generate have very less actions or verbal things going on and most of the time gives me some kind of summary with no dialogues. In short I like things slow( if you know what I mean ).
But still I must say that my go to model for story telling was Guanaco GGML model (as I lack a good gpu). And so far it also gave me pretty good result. But now I can a ton of new models have been released with higher context.
I had recently posted this question in the oobabooga community and I was intrigued by the contribution of two people so I would like to share it here . The summary of their workflow is as follows.
# Getting started:
You need to give the model something to work from . So you will be giving it a few paragraphs of the beginning of the story which will guide the model with the style of literature, dialogue style and lot characteristics and premise of the story. But who has type to write all that, you can again use the and AI model just for this. For example ask it describe the city of Venice and focus on two people .You can also add other things to this beginning to get the result you want.
# Moving on :
The next step is feed the above created paragraphs to the notebook tab in oobabooga ( or respective fields in different GUI). In the notebook you will add in your own sentences at the end of para so that the model knows this is where it has to start adding. At this place you will be trying to divert the story to direction you want it to flow. To do that you will add phrases like ".....but suddenly there was loud scream heard from the...." the model will take this and start typing from "heard from the....".
From here on out you have to guide the story at each instances (for which even I am still learning to get the best result.)
# Multiple Models
One other fascinating thing that I got to know from the people in the previous post is the use of multiple models . Up until 2k context use a 33B model such as Vicuna, then use a Llama 2 model such which has 4k context. and further after that use large context models. But in order for you need to set the settings right .
In oobabooga you need to set the n\_ctx (Context) and compress\_pos\_emb (Context/2048) correctly depending on the model used . (there are actually other settings as well which I am also still figuring out.
This is just scratching the surface and I am only beginning to understand the power of AI.
So I like to invite the community to share your methods, models you use , prompts , settings or your workflow to get good result so that it can help the rest of us as well.
PS: Sorry if things are a bit all over the place , I am still learning myself. And thanks to culturedonichan,Grapecharacter and bankok padang for helping me out on all this things.
| 2023-09-04T12:33:07 |
https://www.reddit.com/r/LocalLLaMA/comments/169qt0o/story_telling_workflow/
|
Darkmeme9
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169qt0o
| false | null |
t3_169qt0o
|
/r/LocalLLaMA/comments/169qt0o/story_telling_workflow/
| false | false |
self
| 1 | null |
Tweaking the llama2 architecture
| 1 |
I want to tweak llama2-7b-hf model to include skip connections and layer norm in between certain layers. Is it possible to do this in pytorch and/or hugginface format?
| 2023-09-04T13:04:02 |
https://www.reddit.com/r/LocalLLaMA/comments/169rglv/tweaking_the_llama2_architecture/
|
Ornery-Young-7346
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169rglv
| false | null |
t3_169rglv
|
/r/LocalLLaMA/comments/169rglv/tweaking_the_llama2_architecture/
| false | false |
self
| 1 | null |
Python LLama 2 on windows
| 1 |
Hello Guys,
I am looking for a way to run llama 2 on windows by python.
I have read that it is done by llama.cpp, but unfortunately it is not running for me...
My error: ModuleNotFoundError: No module named 'llama\_cpp'
But llama cpp is already installed
Is using llama\_cpp the only way to run llama 2?
I need to run it on python since I would like to generate Q&A from text
I also appreciate any other methods!
Thanks!
| 2023-09-04T13:23:00 |
https://www.reddit.com/r/LocalLLaMA/comments/169rwof/python_llama_2_on_windows/
|
Enkay55
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169rwof
| false | null |
t3_169rwof
|
/r/LocalLLaMA/comments/169rwof/python_llama_2_on_windows/
| false | false |
self
| 1 | null |
OTS Structured Document Segmentation Suggestions for RAG Pipeline
| 1 |
[removed]
| 2023-09-04T13:24:57 |
https://www.reddit.com/r/LocalLLaMA/comments/169ry7h/ots_structured_document_segmentation_suggestions/
|
RedditAppSucksDicks
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169ry7h
| false | null |
t3_169ry7h
|
/r/LocalLLaMA/comments/169ry7h/ots_structured_document_segmentation_suggestions/
| false | false |
default
| 1 | null |
Question: How do I post a job for someone with experience fine tuning and configuring a model?
| 1 |
We've got a couple of projects from clients that we originally pitched using gpt/palm, but I feel that we could we getting better performance (both cost and quality) with a finetune. In addition a potential client has a policy banning any model not on their own Infra and we are fully booked.
So I'm looking to hire someone that has experimented enough on these models to know what they are doing, but the market is so young I don't even know what to call the role of where to post it 😅
Any ideas?
(also of course, open to applications through DM)
| 2023-09-04T13:36:31 |
https://www.reddit.com/r/LocalLLaMA/comments/169s7vs/question_how_do_i_post_a_job_for_someone_with/
|
alexmrv
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169s7vs
| false | null |
t3_169s7vs
|
/r/LocalLLaMA/comments/169s7vs/question_how_do_i_post_a_job_for_someone_with/
| false | false |
self
| 1 | null |
WizardLM is writing replies from "User:" After it's Own Response!
| 1 |
Decided to try and run WizardLM "ehartford/WizardLM-7B-Uncensored" for fun - to see if I could get it to help me with a reverse engineering project that I am working on. Normal models don't help as they think I am a hacker.
The model loads, and I can prompt it, but many times, after it's response, it continues on with "User: Thanks for the great info" or something similar. It seems to be confused and thinking this is part of the answer.
Have I set this up wrong? Or is this a training data issue? I am using the model instruction template Vicuna-v1.1, as that is what the text gen web-ui recommended.
Thanks!
| 2023-09-04T13:56:27 |
https://www.reddit.com/r/LocalLLaMA/comments/169soe5/wizardlm_is_writing_replies_from_user_after_its/
|
JustinPooDough
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169soe5
| false | null |
t3_169soe5
|
/r/LocalLLaMA/comments/169soe5/wizardlm_is_writing_replies_from_user_after_its/
| false | false |
self
| 1 | null |
WizardLM is writing replies from "User:" After it's Own Response!
| 1 |
Decided to try and run WizardLM "ehartford/WizardLM-7B-Uncensored" for fun - to see if I could get it to help me with a reverse engineering project that I am working on. Normal models don't help as they think I am a hacker.
The model loads, and I can prompt it, but many times, after it's response, it continues on with "User: Thanks for the great info" or something similar. It seems to be confused and thinking this is part of the answer.
Have I set this up wrong? Or is this a training data issue? I am using the model instruction template Vicuna-v1.1, as that is what the text gen web-ui recommended.
Thanks!
| 2023-09-04T13:56:27 |
https://www.reddit.com/r/LocalLLaMA/comments/169soec/wizardlm_is_writing_replies_from_user_after_its/
|
JustinPooDough
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169soec
| false | null |
t3_169soec
|
/r/LocalLLaMA/comments/169soec/wizardlm_is_writing_replies_from_user_after_its/
| false | false |
self
| 1 | null |
Best plan on runpod.ai for running a model for rpg/storywriting (like mythomax)
| 1 |
I wanna run a really model on the cloud to use it for a rpg/solo play. I am wanna use a model with a high context memory and some nice parameters. I looked up runpod and the I'm not sure what plan to choose for the same (I do not have enough resources on my local). Can someone tell me which plan I should be using? or is there a better alternative to runpod for my task? I do not intend to use stable-diffusion. Only a text model.
| 2023-09-04T14:17:58 |
https://www.reddit.com/r/LocalLLaMA/comments/169t7g5/best_plan_on_runpodai_for_running_a_model_for/
|
Mani_and_5_others
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
169t7g5
| false | null |
t3_169t7g5
|
/r/LocalLLaMA/comments/169t7g5/best_plan_on_runpodai_for_running_a_model_for/
| false | false |
self
| 1 | null |
Running llama 2(any model) in nodejs?
| 1 |
I'm trying to run llama2 model by using node js but am constantly running into errors. The langchain documentation recommends some other npm package and there are some tutorials out there which are using a totally different npm package. I've tried everything(even moved the file to D) in order to make the path simple, but yet the same error pops up.
Can anybody tell me how to correctly set up local llm with nodejs? Fed up of errors and I've been trying since yesterday!
Please help!!
| 2023-09-04T14:58:25 |
ContributionFun3037
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
169u7uy
| false | null |
t3_169u7uy
|
/r/LocalLLaMA/comments/169u7uy/running_llama_2any_model_in_nodejs/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'iRLT8rdRwC5gQtd-UvcLOZxBTqino1NRhn9RgH-qsZg', 'resolutions': [{'height': 58, 'url': 'https://preview.redd.it/lbsexelh69mb1.jpg?width=108&crop=smart&auto=webp&s=da11e411f7bf367fe945f463eeb9eacc874668ba', 'width': 108}, {'height': 116, 'url': 'https://preview.redd.it/lbsexelh69mb1.jpg?width=216&crop=smart&auto=webp&s=6583b6cfd53967481313f84e1f071e62cfa45dc4', 'width': 216}, {'height': 172, 'url': 'https://preview.redd.it/lbsexelh69mb1.jpg?width=320&crop=smart&auto=webp&s=ef6e68e13d8d70ccca29abc11983bc4ba061900a', 'width': 320}, {'height': 344, 'url': 'https://preview.redd.it/lbsexelh69mb1.jpg?width=640&crop=smart&auto=webp&s=21798f8da25be9d19e5932960cf819ba54289089', 'width': 640}, {'height': 516, 'url': 'https://preview.redd.it/lbsexelh69mb1.jpg?width=960&crop=smart&auto=webp&s=963a1fcc2cb42f2c8984632a6dc45e3e5b094a70', 'width': 960}, {'height': 580, 'url': 'https://preview.redd.it/lbsexelh69mb1.jpg?width=1080&crop=smart&auto=webp&s=b4f443e9559e5ffc98d1dc8c48e7f367e662be01', 'width': 1080}], 'source': {'height': 860, 'url': 'https://preview.redd.it/lbsexelh69mb1.jpg?auto=webp&s=c6fc7088048a1290163ec0444f459c6db4b152ff', 'width': 1600}, 'variants': {}}]}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.