CoRNStack: High-Quality Contrastive Data for Better Code Ranking
Paper
•
2412.01007
•
Published
Hugging Face Party @ PyTorch Conference
community
AI & ML interests
None defined yet.
Recent Activity
View all activity
HF-Party's activity
bstadt
authored
2
papers
about 10 hours ago
Post
926
DeepGit: Your GitHub Gold Digger! 💰🚀
Hey Hugging Face gang! Meet DeepGit—my open-source sidekick that rips through GitHub to snag repos that fit you. Done with dead-end searches? Me too. Built it with LangGraph and some dope tricks:
Embeddings grab the good stuff (HF magic, baby!)
Re-ranking nails the best picks
Snoops docs, code, and buzz in one slick flow
Drops a clean list of hidden gems 💎
Unearth that sneaky ML lib or Python gem—run python app.py or langgraph dev and boom! Peek it at https://github.com/zamalali/DeepGit. Fork it, tweak it, love it—Docker’s in, HF vibes are strong. Drop a 🌟 or a crazy idea—I’m pumped to jam with you all! 🪂
Hey Hugging Face gang! Meet DeepGit—my open-source sidekick that rips through GitHub to snag repos that fit you. Done with dead-end searches? Me too. Built it with LangGraph and some dope tricks:
Embeddings grab the good stuff (HF magic, baby!)
Re-ranking nails the best picks
Snoops docs, code, and buzz in one slick flow
Drops a clean list of hidden gems 💎
Unearth that sneaky ML lib or Python gem—run python app.py or langgraph dev and boom! Peek it at https://github.com/zamalali/DeepGit. Fork it, tweak it, love it—Docker’s in, HF vibes are strong. Drop a 🌟 or a crazy idea—I’m pumped to jam with you all! 🪂
Post
841
🚀 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)
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)
Aurelien-Morgan
posted
an
update
3 days ago
osanseviero
authored
a
paper
4 days ago
csabakecskemeti
posted
an
update
9 days ago
Post
3321
I'm collecting llama-bench results for inference with a llama 3.1 8B q4 and q8 reference models on varoius GPUs. The results are average of 5 executions.
The system varies (different motherboard and CPU ... but that probably that has little effect on the inference performance).
https://devquasar.com/gpu-gguf-inference-comparison/
the exact models user are in the page
I'd welcome results from other GPUs is you have access do anything else you've need in the post. Hopefully this is useful information everyone.
The system varies (different motherboard and CPU ... but that probably that has little effect on the inference performance).
https://devquasar.com/gpu-gguf-inference-comparison/
the exact models user are in the page
I'd welcome results from other GPUs is you have access do anything else you've need in the post. Hopefully this is useful information everyone.
csabakecskemeti
posted
an
update
11 days ago
Post
2351
Managed to get my hands on a 5090FE, it's beefy
| llama 8B Q8_0 | 7.95 GiB | 8.03 B | CUDA | 99 | pp512 | 12207.44 ± 481.67 |
| llama 8B Q8_0 | 7.95 GiB | 8.03 B | CUDA | 99 | tg128 | 143.18 ± 0.18 |
Comparison with others GPUs
http://devquasar.com/gpu-gguf-inference-comparison/
| llama 8B Q8_0 | 7.95 GiB | 8.03 B | CUDA | 99 | pp512 | 12207.44 ± 481.67 |
| llama 8B Q8_0 | 7.95 GiB | 8.03 B | CUDA | 99 | tg128 | 143.18 ± 0.18 |
Comparison with others GPUs
http://devquasar.com/gpu-gguf-inference-comparison/
csabakecskemeti
posted
an
update
14 days ago
Post
1796
GTC new model announcement now from Nvidia
nvidia/Llama-3_3-Nemotron-Super-49B-v1
GGUFs:
DevQuasar/nvidia.Llama-3_3-Nemotron-Super-49B-v1-GGUF
Enjoy!
nvidia/Llama-3_3-Nemotron-Super-49B-v1
GGUFs:
DevQuasar/nvidia.Llama-3_3-Nemotron-Super-49B-v1-GGUF
Enjoy!
Post
4687
smolagents now support vLLM! 🥳
As one of the most popular local inference solutions, the community had been asking us to integrate vLLM: after a heavy refactoring of our LLM classes, we've just released smolagents 1.11.0, with a brand new VLLMModel class.
Go try it and tell us what you think!
https://github.com/huggingface/smolagents/blob/45b2c86857b7f7657daaa74e4d17d347e9e2c4a4/src/smolagents/models.py#L497
As one of the most popular local inference solutions, the community had been asking us to integrate vLLM: after a heavy refactoring of our LLM classes, we've just released smolagents 1.11.0, with a brand new VLLMModel class.
Go try it and tell us what you think!
https://github.com/huggingface/smolagents/blob/45b2c86857b7f7657daaa74e4d17d347e9e2c4a4/src/smolagents/models.py#L497
Skylion007
authored
a
paper
18 days ago
csabakecskemeti
posted
an
update
19 days ago
Post
564
Cohere Command-a Q2 quant
DevQuasar/CohereForAI.c4ai-command-a-03-2025-GGUF
6.7t/s on a 3gpu setup (4080 + 2x3090)
(q3, q4 currently uploading)
DevQuasar/CohereForAI.c4ai-command-a-03-2025-GGUF
6.7t/s on a 3gpu setup (4080 + 2x3090)
(q3, q4 currently uploading)
Post
991
Our new Agentic leaderboard is now live!💥
If you ever asked which LLM is best for powering agents, we've just made a leaderboard that ranks them all! Built with @albertvillanova , this ranks LLMs powering a smolagents CodeAgent on subsets of various benchmarks. ✅
🏆 GPT-4.5 comes on top, even beating reasoning models like DeepSeek-R1 or o1. And Claude-3.7-Sonnet is a close second!
The leaderboard also allows you to show the scores of vanilla LLMs (without any agentic setup) on the same benchmarks: this shows the huge improvements brought by agentic setups. 💪
(Note that results will be added manually, so the leaderboard might not always have the latest LLMs)
If you ever asked which LLM is best for powering agents, we've just made a leaderboard that ranks them all! Built with @albertvillanova , this ranks LLMs powering a smolagents CodeAgent on subsets of various benchmarks. ✅
🏆 GPT-4.5 comes on top, even beating reasoning models like DeepSeek-R1 or o1. And Claude-3.7-Sonnet is a close second!
The leaderboard also allows you to show the scores of vanilla LLMs (without any agentic setup) on the same benchmarks: this shows the huge improvements brought by agentic setups. 💪
(Note that results will be added manually, so the leaderboard might not always have the latest LLMs)
- 1 reply
csabakecskemeti
posted
an
update
25 days ago
Post
822
Fine tuning on the edge. Pushing the MI100 to it's limits.
QWQ-32B 4bit QLORA fine tuning
VRAM usage 31.498G/31.984G :D
QWQ-32B 4bit QLORA fine tuning
VRAM usage 31.498G/31.984G :D
-
- 4 replies
Post
1984
🚀 ftBoost is LIVE – Stop Struggling with Fine-Tuning Data!
Alright folks, if you’re tired of manually crafting fine-tuning datasets, ftBoost is here to do the heavy lifting. One-click, LangChain-Groq-powered data augmentation that scales your training data in OpenAI, Gemini, Mistral, and LLaMA formats—automatically.
🔥 What’s inside?
✅ Smart Augmentations – Paraphrasing, back translation, synonym swapping & synthetic noise.
✅ No more JSONL headaches – Auto-formats everything for OpenAI, Gemini, Mistral & LLaMA.
✅ Custom tuning – Adjust similarity, diversity, and fluency in real-time.
✅ Upload, generate, download – That’s it.
⚡ If you’re fine-tuning LLMs, this will save you hours.
🚀 Try it now: 👉 zamal/Finetune-Boost
🌟 Give us a star on GitHub!
Let me know what you think & how it boosts your workflow! 🔥
Alright folks, if you’re tired of manually crafting fine-tuning datasets, ftBoost is here to do the heavy lifting. One-click, LangChain-Groq-powered data augmentation that scales your training data in OpenAI, Gemini, Mistral, and LLaMA formats—automatically.
🔥 What’s inside?
✅ Smart Augmentations – Paraphrasing, back translation, synonym swapping & synthetic noise.
✅ No more JSONL headaches – Auto-formats everything for OpenAI, Gemini, Mistral & LLaMA.
✅ Custom tuning – Adjust similarity, diversity, and fluency in real-time.
✅ Upload, generate, download – That’s it.
⚡ If you’re fine-tuning LLMs, this will save you hours.
🚀 Try it now: 👉 zamal/Finetune-Boost
🌟 Give us a star on GitHub!
Let me know what you think & how it boosts your workflow! 🔥
csabakecskemeti
posted
an
update
about 1 month ago
Post
1954
-UPDATED-
4bit inference is working! The blogpost is updated with code snippet and requirements.txt
https://devquasar.com/uncategorized/all-about-amd-and-rocm/
-UPDATED-
I've played around with an MI100 and ROCm and collected my experience in a blogpost:
https://devquasar.com/uncategorized/all-about-amd-and-rocm/
Unfortunately I've could not make inference or training work with model loaded in 8bit or use BnB, but did everything else and documented my findings.
4bit inference is working! The blogpost is updated with code snippet and requirements.txt
https://devquasar.com/uncategorized/all-about-amd-and-rocm/
-UPDATED-
I've played around with an MI100 and ROCm and collected my experience in a blogpost:
https://devquasar.com/uncategorized/all-about-amd-and-rocm/
Unfortunately I've could not make inference or training work with model loaded in 8bit or use BnB, but did everything else and documented my findings.
-
- 4 replies
Post
4797
We now have a Deep Research for academia: SurveyX automatically writes academic surveys nearly indistinguishable from human-written ones 🔥
Researchers from Beijing and Shanghai just published the first application of a deep research system to academia: their algorithm, given a question, can give you a survey of all papers on the subject.
To make a research survey, you generally follow two steps, preparation (collect and organize papers) and writing (outline creation, writing, polishing). Researchers followed the same two steps and automated them.
🎯 For the preparation part, a key part is find all the important references on the given subject.
Researchers first cast a wide net of all relevant papers. But then finding the really important ones is like distilling knowledge from a haystack of information. To solve this challenge, they built an “AttributeTree” object that structures key information from citations. Ablating these AttributeTrees significantly decreased structure and synthesis scores, so they were really useful!
📝 For the writing part, key was to get a synthesis that's both short and true. This is not easy to get with LLMs! So they used methods like LLM-based deduplication to shorten the too verbose listings made by LLMs, and RAG to grab original quotes instead of made-up ones.
As a result, their system outperforms previous approaches by far!
As assessed by LLM-judges, the quality score os SurveyX even approaches this of human experts, with 4.59/5 vs 4.75/5 🏆
I advise you to read the paper, it's a great overview of the kind of assistants that we'll get in the short future! 👉 SurveyX: Academic Survey Automation via Large Language Models (2502.14776)
Their website shows examples of generated surveys 👉 http://www.surveyx.cn/
Researchers from Beijing and Shanghai just published the first application of a deep research system to academia: their algorithm, given a question, can give you a survey of all papers on the subject.
To make a research survey, you generally follow two steps, preparation (collect and organize papers) and writing (outline creation, writing, polishing). Researchers followed the same two steps and automated them.
🎯 For the preparation part, a key part is find all the important references on the given subject.
Researchers first cast a wide net of all relevant papers. But then finding the really important ones is like distilling knowledge from a haystack of information. To solve this challenge, they built an “AttributeTree” object that structures key information from citations. Ablating these AttributeTrees significantly decreased structure and synthesis scores, so they were really useful!
📝 For the writing part, key was to get a synthesis that's both short and true. This is not easy to get with LLMs! So they used methods like LLM-based deduplication to shorten the too verbose listings made by LLMs, and RAG to grab original quotes instead of made-up ones.
As a result, their system outperforms previous approaches by far!
As assessed by LLM-judges, the quality score os SurveyX even approaches this of human experts, with 4.59/5 vs 4.75/5 🏆
I advise you to read the paper, it's a great overview of the kind of assistants that we'll get in the short future! 👉 SurveyX: Academic Survey Automation via Large Language Models (2502.14776)
Their website shows examples of generated surveys 👉 http://www.surveyx.cn/
csabakecskemeti
posted
an
update
about 1 month ago
Post
2776
Testing Training on AMD/ROCm the first time!
I've got my hands on an AMD Instinct MI100. It's about the same price used as a V100 but on paper has more TOPS (V100 14TOPS vs MI100 23TOPS) also the HBM has faster clock so the memory bandwidth is 1.2TB/s.
For quantized inference it's a beast (MI50 was also surprisingly fast)
For LORA training with this quick test I could not make the bnb config works so I'm running the FT on the fill size model.
Will share all the install, setup and setting I've learned in a blog post, together with the cooling shroud 3D design.
I've got my hands on an AMD Instinct MI100. It's about the same price used as a V100 but on paper has more TOPS (V100 14TOPS vs MI100 23TOPS) also the HBM has faster clock so the memory bandwidth is 1.2TB/s.
For quantized inference it's a beast (MI50 was also surprisingly fast)
For LORA training with this quick test I could not make the bnb config works so I'm running the FT on the fill size model.
Will share all the install, setup and setting I've learned in a blog post, together with the cooling shroud 3D design.
-
- 8 replies
Post
6001
SmolVLM-2 and SigLIP-2 are now part of
They're added on top of the v4.49.0 release, and can be installed from the following tags:
This marks a new beginning for the release process of transformers. For the past five years, we've been doing monthly releases featuring many models (v4.49.0, the latest release, features 9 new architectures).
Starting with SmolVLM-2 & SigLIP2, we'll now additionally release tags supporting new models on a stable branch. These models are therefore directly available for use by installing from the tag itself. These tags will continue to be updated with fixes applied to these models.
Going forward, continue expecting software releases following semantic versioning: v4.50.0 will have ~10 new architectures compared to v4.49.0, as well as a myriad of new features, improvements and bug fixes. Accompanying these software releases, we'll release tags offering brand new models as fast as possible, to make them accessible to all immediately.
transformers in dedicated releases!They're added on top of the v4.49.0 release, and can be installed from the following tags:
v4.49.0-SmolVLM-2 and v4.49.0-SigLIP-2.This marks a new beginning for the release process of transformers. For the past five years, we've been doing monthly releases featuring many models (v4.49.0, the latest release, features 9 new architectures).
Starting with SmolVLM-2 & SigLIP2, we'll now additionally release tags supporting new models on a stable branch. These models are therefore directly available for use by installing from the tag itself. These tags will continue to be updated with fixes applied to these models.
Going forward, continue expecting software releases following semantic versioning: v4.50.0 will have ~10 new architectures compared to v4.49.0, as well as a myriad of new features, improvements and bug fixes. Accompanying these software releases, we'll release tags offering brand new models as fast as possible, to make them accessible to all immediately.
- 1 reply
Post
3079
Less is More for Reasoning (LIMO): a 32B model fine-tuned with 817 examples can beat o1-preview on math reasoning! 🤯
Do we really need o1's huge RL procedure to see reasoning emerge? It seems not.
Researchers from Shanghai Jiaotong University just demonstrated that carefully selected examples can boost math performance in large language models using SFT —no huge datasets or RL procedures needed.
Their procedure allows Qwen2.5-32B-Instruct to jump from 6.5% to 57% on AIME and from 59% to 95% on MATH, while using only 1% of the data in previous approaches.
⚡ The Less-is-More Reasoning Hypothesis:
‣ Minimal but precise examples that showcase optimal reasoning patterns matter more than sheer quantity
‣ Pre-training knowledge plus sufficient computational resources at inference levels up math skills
➡️ Core techniques:
‣ High-quality reasoning chains with self-verification steps
‣ 817 handpicked problems that encourage deeper reasoning
‣ Enough inference-time computation to allow extended reasoning
💪 Efficiency gains:
‣ Only 817 examples instead of 100k+
‣ 40.5% absolute improvement across 10 diverse benchmarks, outperforming models trained on 100x more data
This really challenges the notion that SFT leads to memorization rather than generalization! And opens up reasoning to GPU-poor researchers 🚀
Read the full paper here 👉 LIMO: Less is More for Reasoning (2502.03387)
Do we really need o1's huge RL procedure to see reasoning emerge? It seems not.
Researchers from Shanghai Jiaotong University just demonstrated that carefully selected examples can boost math performance in large language models using SFT —no huge datasets or RL procedures needed.
Their procedure allows Qwen2.5-32B-Instruct to jump from 6.5% to 57% on AIME and from 59% to 95% on MATH, while using only 1% of the data in previous approaches.
⚡ The Less-is-More Reasoning Hypothesis:
‣ Minimal but precise examples that showcase optimal reasoning patterns matter more than sheer quantity
‣ Pre-training knowledge plus sufficient computational resources at inference levels up math skills
➡️ Core techniques:
‣ High-quality reasoning chains with self-verification steps
‣ 817 handpicked problems that encourage deeper reasoning
‣ Enough inference-time computation to allow extended reasoning
💪 Efficiency gains:
‣ Only 817 examples instead of 100k+
‣ 40.5% absolute improvement across 10 diverse benchmarks, outperforming models trained on 100x more data
This really challenges the notion that SFT leads to memorization rather than generalization! And opens up reasoning to GPU-poor researchers 🚀
Read the full paper here 👉 LIMO: Less is More for Reasoning (2502.03387)