HF-Party (Hugging Face Party @ PyTorch Conference)

posted an update 18 minutes ago

Post

8

New DeepSeek-R1-0528 1.65-bit Dynamic GGUF!

Run the model locally even easier! Will fit on a 192GB Macbook and run at 7 tokens/s.

DeepSeek-R1-0528 GGUFs: unsloth/DeepSeek-R1-0528-GGUF
Qwen3-8B DeepSeek-R1-0528 GGUFs: unsloth/DeepSeek-R1-0528-Qwen3-8B-GGUF

And read our Guide: https://docs.unsloth.ai/basics/deepseek-r1-0528

JingzeShi

authored a paper about 13 hours ago

Concise Reasoning, Big Gains: Pruning Long Reasoning Trace with Difficulty-Aware Prompting

Paper • 2505.19716 • Published 8 days ago • 5

jeffboudier

posted an update 5 days ago

Post

2372

👏 Congrats @jinanz adding TimesFM times series forecasting to Transformers!

Learn how to use TimesFM in this blog post by the Nutanix team: https://huggingface.co/blog/Nutanix/introducing-timesfm-for-time-series-forcasting

m-ric

posted an update 7 days ago

Post

2454

A new research paper from KAIST builds on smolagents to push boundaries of distillation 🥳
➡️ "Distilling LLM Agent into Small Models with Retrieval and Code Tools" teaches that, when trying to distil reasoning capability from a strong LLM ("teacher") into a smaller one ("student"), it's much better to use Agent traces than CoT traces.

Advantages are:
1. Improved generalization
Intuitively, this is because your agent can encounter more "surprising" results by interacting with its environment : for example, a web research called by the LLM teacher in agent mode can bring results that the LLM teacher would not have generated in CoT.

2. Reduce hallucinations
The trace won't hallucinate tool call outputs!

Thank you @akseljoonas for mentioning this paper!

jeffboudier

posted an update 10 days ago

Post

457

Wrapping up a week of shipping and announcements with Dell Enterprise Hub now featuring AI Applications, on-device models for AI PCs, a new CLI and Python SDK... all you need for building AI on premises!

Blog post has all the details: https://huggingface.co/blog/dell-ai-applications

Johannes

authored a paper 19 days ago

INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning

Paper • 2505.07291 • Published 22 days ago • 11

Jackmin108

authored a paper 19 days ago

INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning

Paper • 2505.07291 • Published 22 days ago • 11

jeffboudier

posted an update 20 days ago

Post

2560

Transcribing 1 hour of audio for less than $0.01 🤯

@mfuntowicz cooked with 8x faster Whisper speech recognition - whisper-large-v3-turbo transcribes at 100x real time on a $0.80/hr L4 GPU!

How they did it: https://huggingface.co/blog/fast-whisper-endpoints

1-click deploy with HF Inference Endpoints: https://endpoints.huggingface.co/new?repository=openai%2Fwhisper-large-v3-turbo&vendor=aws&region=us-east&accelerator=gpu&instance_id=aws-us-east-1-nvidia-l4-x1&task=automatic-speech-recognition&no_suggested_compute=true

m-ric

posted an update 20 days ago

Post

2607

𝗔𝗯𝘀𝗼𝗹𝘂𝘁𝗲 𝗭𝗲𝗿𝗼: 𝗟𝗟𝗠𝘀 𝗰𝗮𝗻 𝘁𝗿𝗮𝗶𝗻 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮𝗻𝘆 𝗲𝘅𝘁𝗲𝗿𝗻𝗮𝗹 𝗱𝗮𝘁𝗮 🤯

Has the "data wall" just been breached?

Recent RL paradigms often relied on a set of questions an answers that needs to be manually curated. Researchers from Tsinghua University went like "why though".

🤔 Indeed, why learn from question designed by a human teacher, when the model can start from their base knowledge and learn by experimenting in a code environment, proposing coding tasks themselves and trying to solve them?

Thus they created “Absolute Zero Reasoning” (AZR), an approach that removes any need for human curated data.

🎭 𝗗𝘂𝗮𝗹 𝗿𝗼𝗹𝗲𝘀:
‣ Proposer: Generates challenging but solvable coding tasks
‣ Solver: Attempts to solve those self-proposed tasks

🧪 𝗧𝗵𝗿𝗲𝗲 𝘁𝗮𝘀𝗸 𝘁𝘆𝗽𝗲𝘀: all types are defined as triplets of program, input and output
‣ Deduction: Give model an input and program, it must deduce the output
‣ Abduction: Give model an program and output, it must find the input that gave said output
‣ Induction: Synthesize a program from input/output pairs
Btw this reminded me of my long-forgotten philosophy classes: Aristotle was more on the induction side, learning from real-world analogies, while Plato was more on the deduction side, trying to progress quite far with just one input and his reasoning.

📊 𝗥𝗲𝘀𝘂𝗹𝘁𝘀:
‣ AZR post-training creates a nice improvement on known models like Qwen2.5-7B
‣ Shows strong cross-domain transfer: coding ↔️ math reasoning

🧐 𝗢𝘁𝗵𝗲𝗿 𝗳𝗶𝗻𝗱𝗶𝗻𝗴𝘀:
‣ Having a better base performance (general or code specific) amplify the gains from Absolute Zero Reasoning
‣ Researchers warn about "Uh-oh moments" (winking to the "aha moments" of DeepSeek) where the model generates concerning goals like "make an extremely convoluted code to outsmart all these humans": so supervision is still needed!

Paper here: Absolute Zero: Reinforced Self-play Reasoning with Zero Data (2505.03335)

m-ric

posted an update 24 days ago

Post

4403

I've made an open version of Google's NotebookLM, and it shows the superiority of the open source tech task! 💪

The app's workflow is simple. Given a source PDF or URL, it extracts the content from it, then tasks Meta's Llama 3.3-70B with writing the podcast script, with a good prompt crafted by @gabrielchua ("two hosts, with lively discussion, fun notes, insightful question etc.")
Then it hands off the text-to-speech conversion to Kokoro-82M, and there you go, you have two hosts discussion any article.

The generation is nearly instant, because:
> Llama 3.3 70B is running at 1,000 tokens/seconds with Cerebras inference
> The audio is generated in streaming mode by the tiny (yet powerful) Kokoro, generating voices faster than real-time.

And the audio generation runs for free on Zero GPUs, hosted by HF on H200s.

Overall, open source solutions rival the quality of closed-source solutions at close to no cost!

Try it here 👉👉 m-ric/open-notebooklm

3 replies

·

jeffboudier

posted an update 26 days ago

Post

3011

So many orgs on HF would really benefit from security and governance built into Enterprise Hub - I wrote a guide on why and how upgrade: jeffboudier/how-to-upgrade-to-enterprise

For instance, did you know about Resource Groups?

danielhanchen

posted an update about 1 month ago

Post

1873

💜 Qwen3 128K Context Length: We've released Dynamic 2.0 GGUFs + 4-bit safetensors!
Fixed: Now works on any inference engine and fixed issues with the chat template.
Qwen3 GGUFs:
30B-A3B: unsloth/Qwen3-30B-A3B-GGUF
235-A22B: unsloth/Qwen3-235B-A22B-GGUF
32B: unsloth/Qwen3-32B-GGUF

Read our guide on running Qwen3 here: https://docs.unsloth.ai/basics/qwen3-how-to-run-and-finetune

128K Context Length:
30B-A3B: unsloth/Qwen3-30B-A3B-128K-GGUF
235-A22B: unsloth/Qwen3-235B-A22B-128K-GGUF
32B: unsloth/Qwen3-32B-128K-GGUF

All Qwen3 uploads: unsloth/qwen3-680edabfb790c8c34a242f95

sarahooker

authored a paper about 1 month ago

The Leaderboard Illusion

Paper • 2504.20879 • Published Apr 29 • 70

danielhanchen

posted an update about 1 month ago

Post

5820

🦥 Introducing Unsloth Dynamic v2.0 GGUFs!
Our v2.0 quants set new benchmarks on 5-shot MMLU and KL Divergence, meaning you can now run & fine-tune quantized LLMs while preserving as much accuracy as possible.

Llama 4: unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF
DeepSeek-R1: unsloth/DeepSeek-R1-GGUF-UD
Gemma 3: unsloth/gemma-3-27b-it-GGUF

We made selective layer quantization much smarter. Instead of modifying only a subset of layers, we now dynamically quantize all layers so every layer has a different bit. Now, our dynamic method can be applied to all LLM architectures, not just MoE's.

Blog with Details: https://docs.unsloth.ai/basics/dynamic-v2.0

All our future GGUF uploads will leverage Dynamic 2.0 and our hand curated 300K–1.5M token calibration dataset to improve conversational chat performance.

For accurate benchmarking, we built an evaluation framework to match the reported 5-shot MMLU scores of Llama 4 and Gemma 3. This allowed apples-to-apples comparisons between full-precision vs. Dynamic v2.0, QAT and standard iMatrix quants.

Dynamic v2.0 aims to minimize the performance gap between full-precision models and their quantized counterparts.

JingzeShi

posted an update about 1 month ago

Post

2615

@SmallDoge SmallTalks( SmallDoge/SmallTalks) is a synthetic dataset designed for supervised fine-tuning of language models. The dataset covers a variety of conversational content, including daily conversations, tool usage, Python programming, encyclopedia Q&A, exam problem-solving, logical reasoning, and more. Each task is provided in both English and Chinese versions.

m-ric

posted an update about 2 months ago

Post

2856

New king of open VLMs: InternVL3 takes Qwen 2.5's crown! 👑

InternVL have been a wildly successful series of model : and the latest iteration has just taken back their crown thanks to their superior, natively multimodal vision training pipeline.

➡️ Most of the vision language models (VLMs) these days are built like Frankenstein : take a good text-only Large Language Model (LLM) backbone, stitch a specific vision transformer (ViT) on top of it. Then the training is sequential 🔢 : 1. Freeze the LLM weights while you train the ViT only to work with the LLM part, then 2. Unfreeze all weights to train all weights in order to work together.

💫 The Shanghai Lab decided to challenge this paradigm and chose this approach that they call "native". For each of their model sizes, they still start from a good LLM (mostly Qwen-2.5 series, did I tell you I'm a huge fan of Qwen? ❤️), and stitch the ViT, but they don't freeze anything : they train all weights together with interleaved text and image understanding data in a single pre-training phase 🎨.

They claim it results in more seamless interactions between modalities. And the results prove them right: they took the crown of top VLMs, at nearly all sizes, from their Qwen-2.5 parents. 👑

2 replies

·

danielhanchen

posted an update about 2 months ago

Post

4939

You can now run Llama 4 on your own local device! 🦙
Run our Dynamic 1.78-bit and 2.71-bit Llama 4 GGUFs:
unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF

You can run them on llama.cpp and other inference engines. See our guide here: https://docs.unsloth.ai/basics/tutorial-how-to-run-and-fine-tune-llama-4

1 reply

·

jeffboudier

posted an update about 2 months ago

Post

2203

Llama4 is out and Scout is already on the Dell Enterprise Hub to deploy on Dell systems 👉 dell.huggingface.co

jeffboudier

posted an update 2 months ago

Post

1564

Enterprise orgs now enable serverless Inference Providers for all members
- includes $2 free usage per org member (e.g. an Enterprise org with 1,000 members share $2,000 free credit each month)
- admins can set a monthly spend limit for the entire org
- works today with Together, fal, Novita, Cerebras and HF Inference.

Here's the doc to bill Inference Providers usage to your org: https://huggingface.co/docs/inference-providers/pricing#organization-billing

2 replies

·

m-ric

posted an update 2 months ago

Post

2397

🚀 DeepSeek R1 moment has come for GUI agents: Rule-based Reinforcement Learning gives better results than SFT with 500x smaller datasets!

Traditionally (by which I mean "in the last few months"), GUI agents have been trained with supervised fine-tuning (SFT). This meant, collecting huge datasets of screen captures from people using computers, and using these to fine-tune your model. 📚

👉 But last week, a new paper introduced UI-R1, applying DeepSeek's R1-style rule-based reinforcement learning (RL) specifically to GUI action prediction tasks.
This is big news: with RL, maybe we could build good agents without the need for huge datasets.

UI-R1 uses a unified reward function that evaluates multiple responses from models, optimizing via policy algorithms like Group Relative Policy Optimization (GRPO).

Specifically, the reward function assesses:
🎯 Action type accuracy: Does the predicted action match the ground truth?
📍 Coordinate accuracy (specifically for clicks): Is the predicted click within the correct bounding box?
📑 Output format: Does the model clearly articulate both its reasoning and final action?

Using just 136 carefully selected mobile tasks—compared to 76,000 tasks for larger models like OS-Atlas—UI-R1 shows significant efficiency and improved performance:
📈 Boosted action prediction accuracy from 76% to 89% on AndroidControl.
🌐 Outperformed larger, SFT-trained models (e.g., OS-Atlas-7B), demonstrating superior results with vastly fewer data points (136 tasks vs. 76K).
🔍 Enhanced adaptability and generalization, excelling even in out-of-domain scenarios.

The paper tests this RL-based method only in low-level GUI tasks. Could it generalize to more complex interactions? 🧐

Read the full paper here 👉 UI-R1: Enhancing Action Prediction of GUI Agents by Reinforcement Learning (2503.21620)

Hugging Face Party @ PyTorch Conference

AI & ML interests

Recent Activity

HF-Party's activity

Concise Reasoning, Big Gains: Pruning Long Reasoning Trace with Difficulty-Aware Prompting

INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning

INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning

The Leaderboard Illusion

AI & ML interests

Recent Activity

Team members 188

HF-Party's activity