File size: 3,983 Bytes
a47eb6a
3f30ef0
e574d26
 
 
 
 
 
a47eb6a
e574d26
 
3f30ef0
e574d26
 
 
49ae096
 
 
 
e574d26
49ae096
e574d26
49ae096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3f30ef0
49ae096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e574d26
 
 
49ae096
e574d26
 
 
49ae096
e574d26
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49ae096
e574d26
3f30ef0
e574d26
 
 
3f30ef0
e574d26
49ae096
 
 
 
 
 
e574d26
 
 
49ae096
e574d26
49ae096
e574d26
49ae096
e574d26
 
 
49ae096
 
 
 
3f30ef0
e574d26
 
 
49ae096
e574d26
 
 
49ae096
e574d26
 
 
49ae096
e574d26
49ae096
e574d26
49ae096
e574d26
 
 
49ae096
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
---
license: apache-2.0
datasets:
- teknium/OpenHermes-2.5
language:
- en
library_name: transformers
pipeline_tag: text-generation
---
# Model Card for neoncortex/mini-mistral-openhermes-2.5-chatml-test

A tiny Mistral model trained as an experiment on teknium/OpenHermes-2.5.

## Model Details

A 63M parameter auto-regressive LM using Mistral architecture as a base.
- Multi-query Attention instead of Grouped-query Attention.
- Sliding window is disabled.
- Modified ChatML instead of Mistral chat template - TL;DR I used '<|im_start|>human' instead of '<|im_start|>user'

### Model Description

Just doing it to see what happens.

It'll take about 40 to 45 hours to train on two Nvidia RTX 3060 12GB.

It uses ChatML for the chat template, but I fucked up the template in the dataset,
using '<|im_start|>human' instead of '<|im_start|>user'. ¯\_(ツ)_/¯
So, here's the bits:

```
{%- set ns = namespace(found=false) -%}
{%- for message in messages -%}
    {%- if message['role'] == 'system' -%}
        {%- set ns.found = true -%}
    {%- endif -%}
{%- endfor -%}
{%- for message in messages %}
    {%- if message['role'] == 'system' -%}
        {{- '<|im_start|>system\n' + message['content'].rstrip() + '<|im_end|>\n' -}}
    {%- else -%}
        {%- if message['role'] == 'human' -%}
            {{-'<|im_start|>human\n' + message['content'].rstrip() + '<|im_end|>\n'-}}
        {%- else -%}
            {{-'<|im_start|>assistant\n' + message['content'] + '<|im_end|>\n' -}}
        {%- endif -%}
    {%- endif -%}
{%- endfor -%}
{%- if add_generation_prompt -%}
    {{-'<|im_start|>assistant\n'-}}
{%- endif -%}
```

- **Developed by:** gronkomatic
- **Funded by:** gronkomatic
- **Shared by:** gronkomatic
- **Model type:** Mistral
- **Language(s) (NLP):** English, maybe others I dunno
- **License:** OpenRAIL, IDGAF

### Model Sources

Exclusively available right here on HuggingFace!

- **Repository:** https://huggingface.co/neoncortex/mini-mistral-openhermes-2.5-chatml-test
- **Paper:** LoL
- **Demo:** Just download it in Oobabooga and use the modified chatML template above. Maybe I'll throw together a Space or something.

## Uses

If you wanna have a laugh at how bad it is then go ahead, but I wouldn't expect much from it.

### Out-of-Scope Use

This model won't work well for pretty much everything, probably.

## How to Get Started with the Model

Use the code below to get started with the model.

[More Information Needed]

## Training Details

### Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

[More Information Needed]

### Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

#### Preprocessing

I took the OpenHermes 2.5 dataset and formatted it with ChatML.

#### Training Hyperparameters

- **Training regime:** bf16 mixed precision

#### Speeds, Sizes, Times

epochs: 9
steps: 140976
batches per device: 6
1.04it/s

## Evaluation

I tried to run evals but the eval suite just laughed at me.

## Model Examination

Don't be rude.

## Environmental Impact

- **Hardware Type:** I already told you. Try and keep up.
- **Hours used:** ~45 x 2 I guess.
- **Cloud Provider:** gronkomatic
- **Compute Region:** myob
- **Carbon Emitted:** Yes, definitely

### Compute Infrastructure

I trained it on my PC with no side on it because I like to watch the GPUs do their work.

#### Hardware

2 x Nvidia RTX 3060 12GB

#### Software

The wonderful free stuff at HuggingFace (https://huggingface.co)[https://huggingface.co]: transformers, datasets, trl

## Model Card Authors

gronkomatic, unless you're offended by something, in which case it was hacked by hackers.

## Model Card Contact

If you want to send me insults come find me on Reddit I guess u/gronkomatic.