RDP LoRA: Geometry-Driven Identification for Parameter-Efficient Adaptation in Large Language Models
Paper • 2604.19321 • Published • 7
ZeroGPU Explorers
community
AI & ML interests
None defined yet.
Recent Activity
View all activity
ozayezerceli
authored a
paper 3 days ago
Why doesn't anyone host llms in zerogpu spaces?
8
#172 opened 10 days ago
by 
Reality123b
Why doesn't anyone host llms in zerogpu spaces?
8
#172 opened 10 days ago
by Reality123b
sergiopaniego
posted an update 11 days ago
Post
1171
Earlier this month, Apple introduced Simple Self-Distillation: a fine-tuning method that improves models on coding tasks just by sampling from the model and training on its own outputs with plain cross-entropy
And… it's already supported in TRL, built by Kashif Rasul. you can really feel the pace of development in the team 🐎
Paper by Ruixiang ZHANG, He Bai, Huangjie Zheng, Navdeep Jaitly, Ronan Collobert, Yizhe Zhang at Apple 🍎
How it works: the model generates completions at a training-time temperature (T_train) with top_k/top_p truncation, then fine-tunes on them with plain cross-entropy. no labels or verifier needed
You can try it right away with this ready-to-run example (Qwen3-4B on rStar-Coder):
https://github.com/huggingface/trl/blob/main/trl/experimental/ssd/ssd.py
or benchmark a checkpoint with the eval script:
https://github.com/huggingface/trl/blob/main/trl/experimental/ssd/ssd_eval.py
One neat insight from the paper: T_train and T_eval compose into an effective T_eff = T_train × T_eval, so a broad band of configs works well. even very noisy samples still help
Want to dig deeper?
Paper: Embarrassingly Simple Self-Distillation Improves Code Generation (2604.01193)
Trainer docs: https://huggingface.co/docs/trl/main/en/ssd_trainer
And… it's already supported in TRL, built by Kashif Rasul. you can really feel the pace of development in the team 🐎
Paper by Ruixiang ZHANG, He Bai, Huangjie Zheng, Navdeep Jaitly, Ronan Collobert, Yizhe Zhang at Apple 🍎
How it works: the model generates completions at a training-time temperature (T_train) with top_k/top_p truncation, then fine-tunes on them with plain cross-entropy. no labels or verifier needed
You can try it right away with this ready-to-run example (Qwen3-4B on rStar-Coder):
https://github.com/huggingface/trl/blob/main/trl/experimental/ssd/ssd.py
or benchmark a checkpoint with the eval script:
https://github.com/huggingface/trl/blob/main/trl/experimental/ssd/ssd_eval.py
One neat insight from the paper: T_train and T_eval compose into an effective T_eff = T_train × T_eval, so a broad band of configs works well. even very noisy samples still help
Want to dig deeper?
Paper: Embarrassingly Simple Self-Distillation Improves Code Generation (2604.01193)
Trainer docs: https://huggingface.co/docs/trl/main/en/ssd_trainer
Post
120
I taught an old dog, or in this case model, new tricks. Meet Galactic Reasoning 1.3B: https://huggingface.co/collections/TitleOS/galactic-reasoning-galactica-with-chain-of-thought. By finetuning Meta's (at the time Facebook) Galactica model against 
glaiveai glaiveai/reasoning-v1-20m. After training for 1000 steps on my poor overworked Tesla P40 for 48 hours, I was able to produce a merged FP16, LoRA and quantization Q8 weights. Check out the readme.md for an example CoT.
sergiopaniego
posted an update 17 days ago
Post
387
Great experience yesterday at PyTorch Conf Europe in Paris 🇫🇷
We (w/ @kashif ) talked about training LLMs through interaction, using trajectories across games, browsers, or simulators
Room was packed, a clear sign of interest in where RL post-training is heading.
sharing the slides! 🤓
https://drive.google.com/file/d/16k7YRnf5EJEo0XjXGlRJ_hVeLoFWKyNP/view?usp=sharing
We (w/ @kashif ) talked about training LLMs through interaction, using trajectories across games, browsers, or simulators
Room was packed, a clear sign of interest in where RL post-training is heading.
sharing the slides! 🤓
https://drive.google.com/file/d/16k7YRnf5EJEo0XjXGlRJ_hVeLoFWKyNP/view?usp=sharing
sergiopaniego
posted an update 24 days ago
Post
2803
Gemma 4 💎 is here and it’s strong!
to celebrate, we’re rolling out in TRL:
> support for multimodal tool responses for environments (OpenEnv)
> an example to train it in CARLA for autonomous driving with image-based tool calls
go check it out 🏎️🏎️
blog: https://huggingface.co/blog/gemma4
script: https://github.com/huggingface/trl/blob/main/examples/scripts/openenv/carla_vlm_gemma.py
to celebrate, we’re rolling out in TRL:
> support for multimodal tool responses for environments (OpenEnv)
> an example to train it in CARLA for autonomous driving with image-based tool calls
go check it out 🏎️🏎️
blog: https://huggingface.co/blog/gemma4
script: https://github.com/huggingface/trl/blob/main/examples/scripts/openenv/carla_vlm_gemma.py
sergiopaniego
posted an update 26 days ago
Post
2030
TRL is officially an adult 🥳
excited to announce TRL v1.0❗️
head to the blog to see how we got here and what’s next for this post-training library, designed to keep pace with the field
https://huggingface.co/blog/trl-v1
excited to announce TRL v1.0❗️
head to the blog to see how we got here and what’s next for this post-training library, designed to keep pace with the field
https://huggingface.co/blog/trl-v1
- 2 replies
MaziyarPanahi
posted an update 26 days ago
Post
1787
Training mRNA Language Models Across 25 Species for $165
We built an end-to-end protein AI pipeline covering structure prediction, sequence design, and codon optimization. After comparing multiple transformer architectures for codon-level language modeling, CodonRoBERTa-large-v2 emerged as the clear winner with a perplexity of 4.10 and a Spearman CAI correlation of 0.40, significantly outperforming ModernBERT. We then scaled to 25 species, trained 4 production models in 55 GPU-hours, and built a species-conditioned system that no other open-source project offers. Complete results, architectural decisions, and runnable code below.
https://huggingface.co/blog/OpenMed/training-mrna-models-25-species
We built an end-to-end protein AI pipeline covering structure prediction, sequence design, and codon optimization. After comparing multiple transformer architectures for codon-level language modeling, CodonRoBERTa-large-v2 emerged as the clear winner with a perplexity of 4.10 and a Spearman CAI correlation of 0.40, significantly outperforming ModernBERT. We then scaled to 25 species, trained 4 production models in 55 GPU-hours, and built a species-conditioned system that no other open-source project offers. Complete results, architectural decisions, and runnable code below.
https://huggingface.co/blog/OpenMed/training-mrna-models-25-species
vinesmsuic
authored a
paper 26 days ago
vinesmsuic
submitted a
paper to Daily Papers 27 days ago
MaziyarPanahi
posted an update about 1 month ago
Post
2241
We annotated 119K medical images with two frontier VLMs (Qwen 3.5, Kimi K2.5), cross-validated at 93% agreement, and produced 110K training records, all for under $500. Fine-tuning 3 small models (2-3B params) improved all benchmarks: best model reaches +15.0% average exact match.
Everything is open-sourced: datasets, adapters, and code.
https://huggingface.co/blog/OpenMed/synthvision
Everything is open-sourced: datasets, adapters, and code.
https://huggingface.co/blog/OpenMed/synthvision
- 2 replies
Post
6747
We should really have a release date range slider on the /models page. Tired of "trending/most downloaded" being the best way to sort and still seeing models from 2023 on the first page just because they're embedded in enterprise pipelines and get downloaded repeatedly. "Recently Created/Recently Updated" don't solve the discovery problem considering the amount of noise to sift through.
Slight caveat: Trending actually does have some recency bias, but it's not strong/precise enough.
Slight caveat: Trending actually does have some recency bias, but it's not strong/precise enough.
- 3 replies
sergiopaniego
posted an update about 1 month ago
Post
793
ICYMI, great blog by @kashif and @stas on Ulysses Sequence Parallelism: train with million-token contexts
on 4×H100s: 12x longer sequences, 3.7x throughput
learn how to integrate it with Accelerate, Transformers, and TRL ⤵️
https://huggingface.co/blog/ulysses-sp
on 4×H100s: 12x longer sequences, 3.7x throughput
learn how to integrate it with Accelerate, Transformers, and TRL ⤵️
https://huggingface.co/blog/ulysses-sp
sergiopaniego
posted an update about 2 months ago
Post
462
We just released a big blog surveying 16 OSS frameworks for async RL training of LLMs!
We're building a new async GRPO trainer for TRL and as first step, we needed to understand how the ecosystem solves this problem today.
The problem: in synchronous RL training, generation dominates wall-clock time. 32K-token rollouts on a 32B model take hours while training GPUs sit completely idle. With reasoning models and agentic RL making rollouts longer and more variable, this only gets worse.
The ecosystem converged on the same fix: separate inference + training onto different GPU pools, rollout buffer, and async weight sync.
We compared 16 frameworks across 7 axes: orchestration, buffer design, weight sync, staleness management, partial rollouts, LoRA, and MoE support.
This survey is step one. The async GRPO trainer for TRL is next!
https://huggingface.co/blog/async-rl-training-landscape
We're building a new async GRPO trainer for TRL and as first step, we needed to understand how the ecosystem solves this problem today.
The problem: in synchronous RL training, generation dominates wall-clock time. 32K-token rollouts on a 32B model take hours while training GPUs sit completely idle. With reasoning models and agentic RL making rollouts longer and more variable, this only gets worse.
The ecosystem converged on the same fix: separate inference + training onto different GPU pools, rollout buffer, and async weight sync.
We compared 16 frameworks across 7 axes: orchestration, buffer design, weight sync, staleness management, partial rollouts, LoRA, and MoE support.
This survey is step one. The async GRPO trainer for TRL is next!
https://huggingface.co/blog/async-rl-training-landscape
sergiopaniego
posted an update about 2 months ago
Post
431
Nemotron 3 Super by @nvidia is here! NVIDIA's hybrid Mamba2/Transformer models are now natively supported in transformers (no trust_remote_code needed)
Fine-tune them with TRL in just a few lines of code. Notebook + script included to get started right away. goooo!
- Notebook: https://colab.research.google.com/github/huggingface/trl/blob/main/examples/notebooks/sft_nemotron_3.ipynb
- Script: https://github.com/huggingface/trl/blob/main/examples/scripts/sft_nemotron_3.py
- Collection with all the models: https://huggingface.co/collections/nvidia/nvidia-nemotron-v3
Fine-tune them with TRL in just a few lines of code. Notebook + script included to get started right away. goooo!
- Notebook: https://colab.research.google.com/github/huggingface/trl/blob/main/examples/notebooks/sft_nemotron_3.ipynb
- Script: https://github.com/huggingface/trl/blob/main/examples/scripts/sft_nemotron_3.py
- Collection with all the models: https://huggingface.co/collections/nvidia/nvidia-nemotron-v3
Post
3272
Scaling Pedagogical Pre-training to 10 Billion Tokens
New blog post exploring what happens when you take optimal data mixing insights and scale up the data generation itself.
We built Sutra, a multi-stage framework for generating pedagogical pre-training data guided by a knowledge graph of ~2,000 concepts across 9 domains. The pipeline includes structured content generation, six-dimension quality evaluation, diversity management across 20 content styles, and a cleaning stage to prevent collapse.
The result is codelion/sutra-10B, a 10.2 billion token pedagogical dataset with rich metadata (domain, complexity, prerequisites, quality scores) on every entry.
We trained codelion/SmolLM2-70M on it for 3 full epochs (30.6B tokens) on a single A10 GPU in ~78 hours.
Key finding: perplexity kept improving across epochs, but benchmark gains plateaued fast. At 70M parameters, the model hits a representational ceiling that more data alone can't break through.
Full writeup with comparisons against 7 other datasets, detailed benchmark breakdowns, and connections to recent work on synthetic data scaling, curriculum learning, and data mixing laws: https://huggingface.co/blog/codelion/scaling-pedagogical-pretraining-10-billion-tokens
All datasets at multiple scales (10M, 100M, 1B, 10B) plus seed concepts and an SFT variant are in the Sutra Pedagogical Datasets collection.
New blog post exploring what happens when you take optimal data mixing insights and scale up the data generation itself.
We built Sutra, a multi-stage framework for generating pedagogical pre-training data guided by a knowledge graph of ~2,000 concepts across 9 domains. The pipeline includes structured content generation, six-dimension quality evaluation, diversity management across 20 content styles, and a cleaning stage to prevent collapse.
The result is codelion/sutra-10B, a 10.2 billion token pedagogical dataset with rich metadata (domain, complexity, prerequisites, quality scores) on every entry.
We trained codelion/SmolLM2-70M on it for 3 full epochs (30.6B tokens) on a single A10 GPU in ~78 hours.
Key finding: perplexity kept improving across epochs, but benchmark gains plateaued fast. At 70M parameters, the model hits a representational ceiling that more data alone can't break through.
Full writeup with comparisons against 7 other datasets, detailed benchmark breakdowns, and connections to recent work on synthetic data scaling, curriculum learning, and data mixing laws: https://huggingface.co/blog/codelion/scaling-pedagogical-pretraining-10-billion-tokens
All datasets at multiple scales (10M, 100M, 1B, 10B) plus seed concepts and an SFT variant are in the Sutra Pedagogical Datasets collection.
- 2 replies
BestWishYsh
authored a
paper about 2 months ago
BestWishYsh
posted an update about 2 months ago
Post
3602
🚀 Introducing Helios: a 14B real-time long-video generation model!
It’s completely wild—faster than 1.3B models and achieves this without using self-forcing. Welcome to the new era of video generation! 😎👇
💻 Code: https://github.com/PKU-YuanGroup/Helios
🏠 Page: https://pku-yuangroup.github.io/Helios-Page
📄 Paper: Helios: Real Real-Time Long Video Generation Model (2603.04379)
🔹 True Single-GPU Extreme Speed ⚡️
No need to rely on traditional workarounds like KV-cache, quantization, sparse/linear attention, or TinyVAE. Helios hits an end-to-end 19.5 FPS on a single H100!
Training is also highly accessible: an 80GB VRAM can fit four 14B models.
🔹 Solving Long-Video "Drift" from the Core 🎥
Tired of visual drift and repetitive loops? We ditched traditional hacks (like error banks, self-forcing, or keyframe sampling).
Instead, our innovative training strategy simulates & eliminates drift directly, keeping minute-long videos incredibly coherent with stunning quality. ✨
🔹 3 Model Variants for Full Coverage 🛠️
With a unified architecture natively supporting T2V, I2V, and V2V, we are open-sourcing 3 flavors:
1️⃣ Base: Single-stage denoising for extreme high-fidelity.
2️⃣ Mid: Pyramid denoising + CFG-Zero for the perfect balance of quality & throughput.
3️⃣ Distilled: Adversarial Distillation (DMD) for ultra-fast, few-step generation.
🔹 Day-0 Ecosystem Ready 🌍
We wanted deployment to be a breeze from the second we launched. Helios drops with comprehensive Day-0 hardware and framework support:
✅ Huawei Ascend-NPU
✅ HuggingFace Diffusers
✅ vLLM-Omni
✅ SGLang-Diffusion
Try it out and let us know what you think!
It’s completely wild—faster than 1.3B models and achieves this without using self-forcing. Welcome to the new era of video generation! 😎👇
💻 Code: https://github.com/PKU-YuanGroup/Helios
🏠 Page: https://pku-yuangroup.github.io/Helios-Page
📄 Paper: Helios: Real Real-Time Long Video Generation Model (2603.04379)
🔹 True Single-GPU Extreme Speed ⚡️
No need to rely on traditional workarounds like KV-cache, quantization, sparse/linear attention, or TinyVAE. Helios hits an end-to-end 19.5 FPS on a single H100!
Training is also highly accessible: an 80GB VRAM can fit four 14B models.
🔹 Solving Long-Video "Drift" from the Core 🎥
Tired of visual drift and repetitive loops? We ditched traditional hacks (like error banks, self-forcing, or keyframe sampling).
Instead, our innovative training strategy simulates & eliminates drift directly, keeping minute-long videos incredibly coherent with stunning quality. ✨
🔹 3 Model Variants for Full Coverage 🛠️
With a unified architecture natively supporting T2V, I2V, and V2V, we are open-sourcing 3 flavors:
1️⃣ Base: Single-stage denoising for extreme high-fidelity.
2️⃣ Mid: Pyramid denoising + CFG-Zero for the perfect balance of quality & throughput.
3️⃣ Distilled: Adversarial Distillation (DMD) for ultra-fast, few-step generation.
🔹 Day-0 Ecosystem Ready 🌍
We wanted deployment to be a breeze from the second we launched. Helios drops with comprehensive Day-0 hardware and framework support:
✅ Huawei Ascend-NPU
✅ HuggingFace Diffusers
✅ vLLM-Omni
✅ SGLang-Diffusion
Try it out and let us know what you think!
- 6 replies
MaziyarPanahi
posted an update about 2 months ago
Post
4844
DNA, mRNA, proteins, AI. I spent the last year going deep into computational biology as an ML engineer. This is Part I of what I found. 🧬
In 2024, AlphaFold won the Nobel Prize in Chemistry.
By 2026, the open-source community had built alternatives that outperform it.
That's the story I find most interesting about protein AI right now. Not just the science (which is incredible), but the speed at which open-source caught up. Multiple teams, independently, reproduced and then exceeded AlphaFold 3's accuracy with permissive licenses. The field went from prediction to generation: we're not just modeling known proteins anymore, we're designing new ones.
I spent months mapping this landscape for ML engineers. What the architectures actually are (spoiler: transformers and diffusion models), which tools to use for what, and which ones you can actually ship commercially.
New post on the Hugging Face blog: https://huggingface.co/blog/MaziyarPanahi/protein-ai-landscape
Hope you all enjoy! 🤗
In 2024, AlphaFold won the Nobel Prize in Chemistry.
By 2026, the open-source community had built alternatives that outperform it.
That's the story I find most interesting about protein AI right now. Not just the science (which is incredible), but the speed at which open-source caught up. Multiple teams, independently, reproduced and then exceeded AlphaFold 3's accuracy with permissive licenses. The field went from prediction to generation: we're not just modeling known proteins anymore, we're designing new ones.
I spent months mapping this landscape for ML engineers. What the architectures actually are (spoiler: transformers and diffusion models), which tools to use for what, and which ones you can actually ship commercially.
New post on the Hugging Face blog: https://huggingface.co/blog/MaziyarPanahi/protein-ai-landscape
Hope you all enjoy! 🤗
- 2 replies