blog-explorers (Blog-explorers)

posted an update 1 day ago

Post

1315

Try Liquid AI's all-new multimodal models: LFM2-VL-1.6B & LFM2-VL-450M! Demo with the Gradio UI and ReportLab support and both models are runnable on T4 GPU!

↗ LFM2-VL-1.6B-LiquidAI : https://github.com/PRITHIVSAKTHIUR/Multimodal-Outpost-Notebooks/blob/main/LFM2-VL-1.6B-LiquidAI/LFM2-VL-1.6B_ReportLab.ipynb

↗ LFM2-VL-450M-LiquidAI : https://github.com/PRITHIVSAKTHIUR/Multimodal-Outpost-Notebooks/blob/main/LFM2-VL-450M-LiquidAI/LFM2-VL-450M_ReportLab.ipynb

.
.
.
To know more about it, visit the multimodal outpost notebooks !!

1 reply

·

ZennyKenny

posted an update 1 day ago

Post

1120

It's just a matter of time before all the data leakage and data scraping associated with building, training, and using AI results in some kind of major scandal.

That's why I think this paper by @spintronic is so important: Privacy-Preserving Tabular Synthetic Data Generation Using TabularARGN (2508.06647)

Glad to know that there are already researchers looking to mitigate and address this risk before the s**t hits the fan.

Tonic

in blog-explorers/README 2 days ago

[Support] Community Articles

🚀 🤝 1

86

#5 opened over 1 year ago by

victor

prithivMLmods

posted an update 5 days ago

Post

4291

On the verge of releasing Poseidon-Reasoning-5M, a dataset built to excel in general thought processes, mathematics, and science across a diverse mixture of domains, I’m also dropping the Gargantua-R1-Compact dataset, a collection of over six million high-quality reasoning QA pair traces. 🤗🚀

✦ Gargantua-R1-Compact : prithivMLmods/Gargantua-R1-Compact

from datasets import load_dataset

dataset = load_dataset("prithivMLmods/Gargantua-R1-Compact", split="train")

Additionally, I’m adding the mini version of Gargantua — the Gargantua-R1-Wee : prithivMLmods/Gargantua-R1-Wee

from datasets import load_dataset

dataset = load_dataset("prithivMLmods/Gargantua-R1-Wee", split="train")

The composition spans 73.93% core mathematical reasoning involving problems, proofs, and computational challenges, 12.11% across diverse scientific domains such as physics, chemistry, biology, and interdisciplinary topics, 11.35% in competitive coding covering algorithms and data structures, 1.37% in academic science focusing on research-level methodology, 0.95% in creative and analytical reasoning through logic puzzles and problem-solving tasks, 0.25% in specialized technical areas like MLOps, LLMs, diffusion models, and CUDA, and 0.06% involving data from graphs and charts converted into structured JSON formats. Designed with both rich contextual depth and formal structural clarity, Gargantua-R1-Compact is an optimal resource for advancing research in symbolic reasoning, interpretability, and high-precision question answering in mathematical domains.

✦ Collection : prithivMLmods/gargantua-r1-mod-6896bfd7834e82b89ad2b38b

To know more about it, visit the dataset card of the respective dataset. !!

prithivMLmods

posted an update 6 days ago

Post

2149

I've added the demo of the openbmb/MiniCPM-V-4 model to the Hugging Face Space:
prithivMLmods/Multimodal-VLM-Thinking

✨ MiniCPM-V 4.0 is the latest efficient model in the MiniCPM-V series. The model is built based on SigLIP2-400M and MiniCPM4-3B, with a total of 4.1B parameters. It inherits the strong single-image, multi-image, and video understanding performance of MiniCPM-V 2.6 with largely improved efficiency.

✨ With only 4.1B parameters, MiniCPM-V 4.0 achieves an average score of 69.0 on OpenCompass, a comprehensive evaluation of 8 popular benchmarks. This performance surpasses GPT-4.1-mini-20250414, MiniCPM-V 2.6 (8.1B parameters, OpenCompass 65.2), and Qwen2.5-VL-3B-Instruct (3.8B parameters, OpenCompass 64.5). It also shows good performance in multi-image and video understanding.

The community GPU grant was given by Hugging Face — special thanks to them. 🤗🚀

To know more about it, visit the model card of the respective model. !!

gmongaras

authored a paper 8 days ago

On the Expressiveness of Softmax Attention: A Recurrent Neural Network Perspective

Paper • 2507.23632 • Published 14 days ago • 6

prithivMLmods

posted an update 9 days ago

Post

4172

Qwen Image – The Latest Image Generation Model🔥

Below are some samples generated using the Qwen Image Diffusion Model. Qwen-Image, a 20B MMDiT model for next-generation text-to-image generation, preserves typographic details, layout coherence, and contextual harmony with stunning accuracy. It is especially strong at creating stunning graphic posters with native text. The model is now open-source. [ 𝚀𝚠𝚎𝚗-𝙸𝚖𝚊𝚐𝚎 : Qwen/Qwen-Image ]

⤷ Try the Qwen Image demo here: prithivMLmods/Qwen-Image-Diffusion

⤷ Qwen-Image Technical Report : Qwen-Image Technical Report (2508.02324)
⤷ Qwen Image [GitHub] : https://github.com/QwenLM/Qwen-Image

Even more impressively, it demonstrates a strong ability to understand images. The model supports a wide range of vision-related tasks such as object detection, semantic segmentation, depth and edge (Canny) estimation, novel view synthesis, and image super-resolution. While each task is technically distinct, they can all be viewed as advanced forms of intelligent image editing driven by deep visual understanding. Collectively, these capabilities position Qwen-Image as more than just a tool for generating appealing visuals, it serves as a versatile foundation model for intelligent visual creation and transformation, seamlessly blending language, layout, and imagery.

Qwen-Image uses a dual-stream MMDiT architecture with a frozen Qwen2.5-VL, VAE encoder, RMSNorm for QK-Norm, LayerNorm elsewhere, and a custom MSRoPE scheme for joint image-text positional encoding.

.
.
.
To know more about it, visit the model card of the respective model. !!

a-r-r-o-w

posted an update 10 days ago

Post

1993

You would've implemented the 3-loop matrix multiplication many times as a ML practitioner, but the naive implementation is terrible for GPU performance. Modern GPUs achieve peak performance through careful memory access patterns and minimizing scheduling overhead.

In naive matmul (MxK . KxN), the computation happens in tiles - both for the output matrix and for how you read chunks from the input matrices. Each thread-block processes one output tile by loading corresponding tiles from input (for sum-reduction across K dimension), performing the computation, then terminating. The GPU launches many thread-blocks and schedules them across available streaming multiprocessors (SMs). When an SM finishes one tile, it gets assigned a new thread-block for the next uncomputed tile. This way, multiple output tiles are computed in parallel across the SMs, but we pay the cost for launching thread-blocks each time a new tile is computed.

Persistent matmul changes this approach. Instead of launching thread-blocks to compute some output tiles, computing the results on SMs in parallel, and repeating until all output tiles are computed, you launch only as many thread-blocks as you have SMs available (typically 80-132 on modern GPUs). These thread-blocks stay alive until all output tiles are computed, looping through multiple tiles sequentially. Each persistent thread-block may handle multiple output tiles.

The key benefit is the reduced thread-block launch latency. This persistence strategy, combined with other optimizations like coalesced memory loads/stores, block-tiling, warp-tiling, warp-specialization, double-buffering, ping-pong scheduling and other tricks, helps achieve peak performance. More on this in the future!

Code snippet for testing: https://gist.github.com/a-r-r-o-w/28339b442d164084506c0967029968a8

(Bonus: Since I've wanted to learn Manim for a while, this was a great opportunity to make a visualization for Naive VS Persistent matmul. Enjoy ✨)

3 replies

·

Abhaykoul

posted an update 12 days ago

Post

3773

🚀 Dhanishtha-2.0-preview-0825 Is Here

The Intermediate Thinking Model just leveled up again.

With sharper reasoning, better tool use, and expanded capabilities, Dhanishtha-2.0-preview-0825 is now live and ready to impress.

🧠 What Makes Dhanishtha Special?
Unlike typical CoT models that only thinks one time, Dhanishtha thinks iteratively:

> Think → Answer → Rethink → Improve → Rethink again if needed.

🔗 Try it now: HelpingAI/Dhanishtha-2.0-preview-0825

🔞 Dhanishtha NSFW Preview

For those exploring more expressive and immersive roleplay scenarios, we’re also releasing:

HelpingAI/Dhanishtha-nsfw
A specialized version tuned for adult-themed interactions and character-driven roleplay.

🔗 Explore it here: HelpingAI/Dhanishtha-nsfw

💬 You can also try all of these live at chat.helpingai.co

4 replies

·

prithivMLmods

posted an update 12 days ago

Post

3147

Introducing Camel-Doc-OCR-080125(v2), a document content-structure retrieval VLM designed for content extraction and summarization. This is the second model in the Camel Doc OCR VLM series, following Camel-Doc-OCR-062825(v1). The new version fixes formal table reconstruction issues in both en and zh language, achieving optimal performance for long-context inferences.🤗🐪

⤷ Camel-Doc-OCR(v2) : prithivMLmods/Camel-Doc-OCR-080125
⤷ Camel-Doc-OCR(v1) : prithivMLmods/Camel-Doc-OCR-062825
⤷ Demo : prithivMLmods/core-OCR

Multimodal Model Collections and Spaces:

➝ Camel-Doc-OCR : prithivMLmods/camel-doc-ocr-080125-688c0c61c5dba648756f31f8
➝ Vision-Language (VLr) : prithivMLmods/vision-language-for-reasoning-vlr-6889b3f45917352b5e3a6f7a
➝ Multimodal Spaces : prithivMLmods/multimodal-implementations-67c9982ea04b39f0608badb0
➝ Multimodal VLMs : prithivMLmods/multimodal-vlms-until-july25-688312e6b840e1e156f13027

.
.
.
To know more about it, visit the model card of the respective model. !!

2 replies

·

prithivMLmods

posted an update 14 days ago

Post

1072

Exciting to bring the explicitly grounded experimental reasoning model, Lumian-VLR-7B-Thinking, built on top of Qwen2.5-VL, featuring reasoning-aware trajectories with enhanced spatial perception. Along with this, we’ve also added a demo for the model while bringing some of the latest and most interesting models available on the hub to make full use of the remaining resources.

✨ Multimodal-VLM-Thinking : prithivMLmods/Multimodal-VLM-Thinking
✨ Multimodal-VLM-OCR : prithivMLmods/Multimodal-VLM-OCR

✦ Models used in these spaces:

✨ Lumian-VLR-7B-Thinking : prithivMLmods/Lumian-VLR-7B-Thinking
✨ Enesidaon-VLR-7B-no-Thinking : prithivMLmods/Enesidaon-VLR-7B-no-Thinking
✨ GLM-4.1V-9B-Thinking : zai-org/GLM-4.1V-9B-Thinking
✨ DREX-062225-exp : prithivMLmods/DREX-062225-exp & more ...

✦ Multimodal Model Collections and Spaces:

✨ Vision-Language (VLr) : prithivMLmods/vision-language-for-reasoning-vlr-6889b3f45917352b5e3a6f7a
✨ Multimodal Spaces : prithivMLmods/multimodal-implementations-67c9982ea04b39f0608badb0
✨ Multimodal VLMs : prithivMLmods/multimodal-vlms-until-july25-688312e6b840e1e156f13027

.
.
.
To know more about it, visit the model card of the respective model. !!

prithivMLmods

posted an update 18 days ago

Post

4817

Explore OCR, Captioning, and Visual Understanding with Cutting-Edge Models on Hugging Face. 🤗🧪

I’ve put together a collection of Google Colab notebooks to experiment with some of the most exciting models available on the Hugging Face Hub focused on OCR, image captioning, and visual understanding tasks. [Image-to-Text] / [Image-Text-to-Text]

> 📖 OCR-ReportLab-Notebooks : prithivMLmods/OCR-ReportLab-Notebooks

These notebooks are built for quick prototyping and run on free T4 GPUs, making them perfect for experimentation, testing ideas, or just exploring what’s possible with modern vision-language models.

Note: The experimental notebooks are compiled with models that fit within the T4 GPU (free-tier) limits. More models along with their notebooks will be added over time.

prithivMLmods

posted an update 20 days ago

Post

2378

Excited to introduce the new experimental model "Qwen2.5-VL-7B-Abliterated-Caption-it", which is performing exceptionally well on image captioning tasks. This variant is specifically tailored for Abliterated Captioning and Uncensored Image Captioning. It is designed to generate highly detailed and descriptive captions across a broad range of visual categories including images with complex, sensitive, or nuanced content while handling varying aspect ratios and resolutions.🧪🤗

✨ Try the demo here : https://huggingface.co/spaces/prithivMLmods/Qwen2.5-VL
✨ Qwen2.5-VL-7B-Abliterated-Caption-it : prithivMLmods/Qwen2.5-VL-7B-Abliterated-Caption-it
✨ Multimodal VLMs : prithivMLmods/multimodal-vlms-until-july25-688312e6b840e1e156f13027
✨ Multimodal Implementations : prithivMLmods/multimodal-implementations-67c9982ea04b39f0608badb0

.
.
.
To know more about it, visit the model card of the respective model. !!

prithivMLmods

posted an update 21 days ago

Post

2375

olmOCR [Allen AI] just got an upgrade! 📈🧑‍🍳

The allenai/olmOCR-7B-0725 — fine-tuned with allenai/olmOCR-mix-0225 on top of Qwen/Qwen2.5-VL-7B-Instruct, pushing the boundaries of OCR technology. It takes a single document image as input, with the longest side resized to 1288 pixels. High-quality, openly available approach to parsing pdfs and other complex documents optical character recognition.

Try the demo here: prithivMLmods/Multimodal-OCR

✨ Model: allenai/olmOCR-7B-0725
✨ Model [fp8]: allenai/olmOCR-7B-0725-FP8
✨ Multimodal Implementations Space Collection: prithivMLmods/multimodal-implementations-67c9982ea04b39f0608badb0

.
.
.
To know more about it, visit the model card of the respective model. !!

ariG23498

posted an update 24 days ago

Post

569

I have always advocated for writing techinical stories without using LLMs.

The following one page editorial really drives the point home.
https://www.nature.com/articles/s44222-025-00323-4

prithivMLmods

posted an update 25 days ago

Post

5116

Upgraded the step-by-step notebook for fine-tuning SigLIP2 on domain-specific image classification tasks. The notebook supports both datasets with predefined train/test splits and those with only a train split, making it suitable for low-resource, custom, and real-world classification scenarios. 📢👉

➺ FineTuning-SigLIP2-Notebook : prithivMLmods/FineTuning-SigLIP2-Notebook

➺ GitHub : https://github.com/PRITHIVSAKTHIUR/FineTuning-SigLIP-2

➺ In the first, datasets include predefined train and test splits, enabling conventional supervised learning and generalization evaluation : prithivMLmods/FineTuning-SigLIP2-Notebook (.ipynb)

➺ In the second scenario, only a training split is available; in such cases, the training set is either partially reserved for validation or reused entirely for evaluation : prithivMLmods/FineTuning-SigLIP2-Notebook (.ipynb)

This flexibility supports experimentation in constrained or domain-specific settings, where standard test annotations may not exist.

prithivMLmods

posted an update 26 days ago

Post

4079

Dropping the general-purpose reasoning dataset Poseidon-Reasoning-5M, which supports general thought processes, math, and science — featuring a diverse mixture of domains 🌊 : prithivMLmods/Poseidon-Reasoning-5M

from datasets import load_dataset

dataset = load_dataset("prithivMLmods/Poseidon-Reasoning-5M", split="data")

The compact version is as follows — Poseidon-Reasoning-Mini-300K : prithivMLmods/Poseidon-Reasoning-Mini-300K

from datasets import load_dataset

dataset = load_dataset("prithivMLmods/Poseidon-Reasoning-Mini-300K", split="train")

Collection : prithivMLmods/poseidon-reasoning-6879ca98e118b307c781a9ba

Abhaykoul

posted an update 29 days ago

Post

3036

🎉 Dhanishtha-2.0-preview-0725 is Now Live

The Intermediate Thinking Model just got even better.
With the new update, Dhanishtha is now sharper, smarter, and trained further on tool use

🧠 What Makes Dhanishtha Different?
Unlike standard COT models that give one-shot responses, Dhanishtha thinks in layers:

> Think → Answer → Rethink → Improve → Rethink again if needed.

HelpingAI/Dhanishtha-2.0-preview-0725