🔄 In a Training Loop

Quentin Gallouédec PRO

qgallouedec

1770 186 147

AI & ML interests

None yet

Recent Activity

updated a bucket about 20 hours ago

qgallouedec/cron

updated a dataset 2 days ago

hf-doc-build/doc-build

updated a bucket 2 days ago

hf-doc-build/doc

View all activity

Organizations

reacted to sergiopaniego's post with 👍 24 days ago

Post

2401

most multi-turn RL loops have a silent bug: you decode the model's output to detect tool calls, then re-tokenize the conversation for the next turn. BPE isn't invertible, so decode then re-encode can land on different ids. gradient ends up on tokens the model never sampled. no crash, just quietly wrong math and broken training

@qgallouedec wrote a super educational blog on MITO (message-in, token-out) vs TITO (token-in, token-out) and how you might fix the problem above

go read it 🤓

https://qgallouedec-tito.hf.space/

reacted to albertvillanova's post with 🤗 2 months ago

Post

4066

🎉 KTO is now part of the stable TRL API

As of Promote KTO to stable API, KTOTrainer and KTOConfig have graduated from trl.experimental to the stable trl API. https://github.com/huggingface/trl/pull/6175

This one closes out a long road. Over the past 6+ months, the "Align KTO with DPO" effort landed ~90 PRs methodically bringing KTO up to the standard we hold for stable trainers, one carefully-scoped change at a time:
- Feature parity with DPO: full VLM support (incl. multi-image), sync_ref_model, PEFT + Liger, ZeRO-3 + PEFT dtype fix, pad_to_multiple_of, activation offloading, IterableDataset and dict eval_dataset, remove_unused_columns, and reference-logprob precomputation at init.
- Consistency with DPO: aligned method order and signatures, tokenization, _prepare_dataset, PEFT handling, ref-model preparation for distributed training, and config layout — plus a new DataCollatorForKTO and output format. Metrics moved into _compute_loss and simplified to direct averages via the shared _metrics attribute.
- Removing legacy baggage: dropped encoder-decoder support, BOS/EOS handling, null_ref_context, generate_during_eval, model_init, preprocess_logits_for_metrics, model/ref adapter names, and several dead config knobs.
- Coverage: a full test suite mirroring DPO, text collator tests, VLM tests, and slow tests.
- The promotion itself: the experimental → stable move (#6175) and shim cleanup (#6287), handled so downstream users get a clean deprecation path.

Honestly, this has been one of the more complex tasks I've taken on since joining the team, not because any single change was hard, but because it demanded sustained consistency across a ~2,000-line trainer, with every branch, comment, and edge case kept in lockstep with DPO.

Huge thanks to everyone who reviewed along the way (especially @qgallouedec ), the incremental review cadence is exactly what kept this maintainable.

KTO now sits on equal footing with our other flagship trainers. 🚀

2 replies

posted an update 4 months ago

Post

10869

Shipped hf-sandbox! 🥡

🧪 Running an eval that executes model-generated C on a few thousand prompts? You probably don't want any of that on your laptop.
Just shipped hf-sandbox, a Modal-style sandbox API on top of Hugging Face Jobs. Spin up an isolated, ephemeral container, run untrusted code, get the result back. No Docker on your laptop, no infra to manage.

Just pip install hf-sandbox.

Early days (v0.1); feedback and issues very welcome:
👉 https://github.com/huggingface/hf-sandbox

1 reply

posted an update 4 months ago

Post

516

**TRL v1.4 is out 🚀** Chunked NLL loss for SFT and a first-class **OpenReward** integration.

**Chunked NLL loss for SFT — drops peak VRAM by up to 14×**

Standard SFT materializes a full [batch × seq × vocab] logits tensor before computing cross-entropy, which dominates peak memory at long context lengths. The new loss_type="chunked_nll" path drops ignored-label tokens before the lm_head matmul and computes cross-entropy in checkpointed chunks of 256.

Peak GPU memory, AdamW fp32:
- Qwen3-14B, 8×H100 FSDP2, 16k seq: 58.9 GB → 38.9 GB
- Qwen3-4B, 1×H100 80GB, 16k seq: OOM → 63.8 GB
- Qwen3-32B, 8×H100 FSDP2, 8k seq: OOM → 71.2 GB

End-to-end it's consistently as fast or faster than nll, and unlocks sequence lengths that don't fit at all under the standard path.

SFTConfig(loss_type="chunked_nll")

Works with PEFT and VLMs out of the box.

**Open Reward Standard environment adapter**

The new trl.experimental.openreward adapter plugs any environment speaking the [Open Reward Standard](https://openrewardstandard.io) protocol into any TRL trainer that takes an environment_factory. One string — a catalog name or a URL — wires the dataset, factory, and reward_func slots; tools are bound dynamically from JSON Schema, no per-env wrapper code:

from trl import GRPOTrainer
from trl.experimental.openreward import OpenRewardSpec

spec = OpenRewardSpec("Eigent/SETA", num_tasks=64)

trainer = GRPOTrainer(
    ...,
    train_dataset=spec.train_dataset,
    environment_factory=spec.environment_factory,
    reward_funcs=spec.reward_funcs,
)

v1.4 also brings MFU helpers for dense + MoE models, GRPO support for Liger 0.8.0 (delta clipping + VESPO + KL bias correction), Tülu 3's length-normalized DPO loss, four more training chat templates (Cohere, Cohere2, Gemma 3, Qwen3-2507), and a 5+ GB CUDA memory leak fix in activation offloading.

Full release notes: https://github.com/huggingface/trl/releases/tag/v1.4.0

posted an update 5 months ago

Post

8251

TRL v1.3 ships day-one training support for Qwen 3.6 🚀

The new Qwen 3.6 family (Qwen/Qwen3.6-27B, Qwen/Qwen3.6-35B-A3B) reuses the Qwen3.5-MoE architecture but ships a slightly different chat template, so we updated the stack end-to-end: new training template with {% generation %} markers, tool-call response schema routing, tiny test models for the VLM matrix.

SFT with assistant-only loss works out of the box:

from trl import SFTConfig, SFTTrainer

trainer = SFTTrainer(
    model="Qwen/Qwen3.6-27B",
    args=SFTConfig(assistant_only_loss=True),
    train_dataset=dataset,
)
trainer.train()

So does GRPO tool-calling — just hand tools=[...] to GRPOTrainer.

v1.3 also brings a new experimental TPO trainer (Triple Preference Optimization), speculative decoding in trl vllm-serve (Qwen3 MTP / Eagle3 drafts), 12 more KTO ↔ DPO alignment PRs (KTO promotion to stable is now in reach), three more {% generation %} chat templates (Gemma/Gemma 2, Phi-3, GLM-4-MoE), and a chunky SFT entropy bug fix.

Full release notes: https://github.com/huggingface/trl/releases/tag/v1.3.0

posted an update 5 months ago

Post

2108

TRL v1.2 introduces the SSDTrainer 🚀

Simple Self-Distillation (SSD) from Apple's paper "Embarrassingly Simple Self-Distillation Improves Code Generation" is now available as an experimental trainer in TRL.

The recipe is as minimal as the name suggests: sample completions from the model itself at a training-time temperature, then fine-tune on those raw, unverified samples with plain cross-entropy. No reward model. No verifier. No teacher model. No reinforcement learning. Just prompts and the model.

from trl.experimental.ssd import SSDConfig, SSDTrainer

trainer = SSDTrainer(
    model="Qwen/Qwen3-4B-Instruct",
    args=SSDConfig(temperature=0.6, top_k=20, top_p=0.95),
    train_dataset=dataset,
)
trainer.train()

v1.2 also ships expanded tool-calling support (LLaMA 3.1 / 3.2, DeepSeek-V3), another round of KTO ↔ DPO alignment getting us closer to promoting KTO to stable, a big GRPO simplification for overlong tool results, deprecation of use_transformers_paged, and key fixes for VLM response parsing.

Full release notes: https://github.com/huggingface/trl/releases/tag/v1.2.0

posted an update 6 months ago

Post

2522

TRL v1.0 is out!

Hugging Face's TRL library is downloaded 3 million times a month. Over 130k models trained with it are public on the Hub, and major projects like @unsloth and @axolotl-ai-co build directly on top of it. v1.0 is the moment we acknowledged that responsibility explicitly, with a real stability contract.

The field hasn't settled. Building stable software in a domain that keeps invalidating its own assumptions is the actual problem we're solving. The answer is a design that can absorb the next shift without breaking what people rely on.

What's in v1.0:
Deep Hugging Face integration, low infrastructure burden
What's next: asynchronous GRPO, better scaling support, and making training legible enough that agents can inspect and steer it.

pip install --upgrade trl

Read more: hf.co/blog/trl-v1

reacted to sergiopaniego's post with 🚀 7 months ago

Post

2561

What happens when you make an LLM drive a car where physics are real and actions can't be undone?

I ported CARLA, the autonomous driving simulator, to OpenEnv and added training support via TRL + Hugging Face Spaces.

The model interacts with the simulator through tool calls (observe, brake, change lane) and learns from a reward signal.

In 50 training steps, Qwen 0.6B learns to swerve and brake to avoid pedestrians in emergency situations.

The project supports text and vision (VLMs can see through a camera sensor), open-world driving with traffic, and multiple driving scenarios.

This builds on the carla-env project by sinatras, which originally placed LLMs inside CARLA for evaluation. We extended it with vision, new scenarios, rubric-based rewards, and made it trainable end-to-end.

Blog: https://huggingface.co/blog/sergiopaniego/bringing-carla-to-openenv-trl/
CARLA env in OpenEnv: https://github.com/meta-pytorch/OpenEnv/tree/main/envs/carla_env
Training script: https://github.com/huggingface/trl/blob/main/examples/scripts/openenv/carla.py

posted an update 7 months ago

Post

3128

@CohereLabs just released 🌿 Tiny Aya: a fully open-source 3B parameter model that speaks 70+ languages 🌍! But there’s a catch:

Tiny Aya is just a language model. It doesn’t support tool calling, the key capability that turns frontier models into powerful *agents*.
So the real question is:

How hard is it to turn Tiny Aya into an agent?

Turns out… it’s simple, thanks to Hugging Face TRL.
We’re sharing a hands-on example showing how to train Tiny Aya to turn it into a tool-calling agent using TRL, unlocking what could become the first *massively multilingual open agent*.

Small model. Global reach. Agent capabilities.

👉 https://github.com/huggingface/trl/blob/main/examples/notebooks/sft_tool_calling.ipynb

1 reply

reacted to sergiopaniego's post with 🔥 8 months ago

Post

2665

New TRL + OpenEnv example! 💥

Fine tune an LLM for playing Sudoku using an RL env via OpenEnv

Includes a script that runs on 1 or multiple GPUs with vLLM, plus a Colab-ready notebook.

Enjoy!

Notebook: https://colab.research.google.com/github/huggingface/trl/blob/main/examples/notebooks/openenv_sudoku_grpo.ipynb

Script: https://github.com/huggingface/trl/blob/main/examples/scripts/openenv/sudoku.py

1 reply

reacted to AtAndDev's post with 🤗 over 1 year ago

Post

3255

deepseek-ai/DeepSeek-R1-0528

This is the end

2 replies

reacted to fdaudens's post with 🔥 over 1 year ago

Post

1829

🔥The Open R1 team just dropped OlympicCoder and it's wild:

- 7B model outperforms Claude 3.7 Sonnet on IOI benchmark (yes, 7B!!)
- 32B crushes all open-weight models tested, even those 100x larger 🤯

Open-sourcing the future of code reasoning! 🚀

Check it out https://huggingface.co/blog/open-r1/update-3

reacted to lewtun's post with 🔥 over 1 year ago

Post

10575

We are reproducing the full DeepSeek R1 data and training pipeline so everybody can use their recipe. Instead of doing it in secret we can do it together in the open!

🧪 Step 1: replicate the R1-Distill models by distilling a high-quality reasoning corpus from DeepSeek-R1.

🧠 Step 2: replicate the pure RL pipeline that DeepSeek used to create R1-Zero. This will involve curating new, large-scale datasets for math, reasoning, and code.

🔥 Step 3: show we can go from base model -> SFT -> RL via multi-stage training.

Follow along: https://github.com/huggingface/open-r1

5 replies

replied to merve's post over 1 year ago

reacted to merve's post with 🔥 over 1 year ago

Post

5666

Oof, what a week! 🥵 So many things have happened, let's recap! merve/jan-24-releases-6793d610774073328eac67a9

Multimodal 💬
- We have released SmolVLM -- tiniest VLMs that come in 256M and 500M, with it's retrieval models ColSmol for multimodal RAG 💗
- UI-TARS are new models by ByteDance to unlock agentic GUI control 🤯 in 2B, 7B and 72B
- Alibaba DAMO lab released VideoLlama3, new video LMs that come in 2B and 7B
- MiniMaxAI released Minimax-VL-01, where decoder is based on MiniMax-Text-01 456B MoE model with long context
- Dataset: Yale released a new benchmark called MMVU
- Dataset: CAIS released Humanity's Last Exam (HLE) a new challenging MM benchmark

LLMs 📖
- DeepSeek-R1 & DeepSeek-R1-Zero: gigantic 660B reasoning models by DeepSeek, and six distilled dense models, on par with o1 with MIT license! 🤯
- Qwen2.5-Math-PRM: new math models by Qwen in 7B and 72B
- NVIDIA released AceMath and AceInstruct, new family of models and their datasets (SFT and reward ones too!)

Audio 🗣️
- Llasa is a new speech synthesis model based on Llama that comes in 1B,3B, and 8B
- TangoFlux is a new audio generation model trained from scratch and aligned with CRPO

Image/Video/3D Generation ⏯️
- Flex.1-alpha is a new 8B pre-trained diffusion model by ostris similar to Flux
- tencent released Hunyuan3D-2, new 3D asset generation from images

7 replies

reacted to lewtun's post with 🔥 over 1 year ago

Post

7126

We outperform Llama 70B with Llama 3B on hard math by scaling test-time compute 🔥

How? By combining step-wise reward models with tree search algorithms :)

We show that smol models can match or exceed the performance of their much larger siblings when given enough "time to think"

We're open sourcing the full recipe and sharing a detailed blog post.

In our blog post we cover:

📈 Compute-optimal scaling: How we implemented DeepMind's recipe to boost the mathematical capabilities of open models at test-time.

🎄 Diverse Verifier Tree Search (DVTS): An unpublished extension we developed to the verifier-guided tree search technique. This simple yet effective method improves diversity and delivers better performance, particularly at large test-time compute budgets.

🧭 Search and Learn: A lightweight toolkit for implementing search strategies with LLMs and built for speed with vLLM

Here's the links:

- Blog post: HuggingFaceH4/blogpost-scaling-test-time-compute

- Code: https://github.com/huggingface/search-and-learn

Enjoy!

2 replies

reacted to cfahlgren1's post with ❤️ almost 2 years ago

Post

3301

You can clean and format datasets entirely in the browser with a few lines of SQL.

In this post, I replicate the process @mlabonne used to clean the new microsoft/orca-agentinstruct-1M-v1 dataset.

The cleaning process consists of:
- Joining the separate splits together / add split column
- Converting string messages into list of structs
- Removing empty system prompts

https://huggingface.co/blog/cfahlgren1/the-beginners-guide-to-cleaning-a-dataset

Here's his new cleaned dataset: mlabonne/orca-agentinstruct-1M-v1-cleaned

1 reply

reacted to Wauplin's post with 🔥 almost 2 years ago

Post

4732

🚀 Exciting News! 🚀

We've just released 𝚑𝚞𝚐𝚐𝚒𝚗𝚐𝚏𝚊𝚌𝚎_𝚑𝚞𝚋 v0.25.0 and it's packed with powerful new features and improvements!

✨ 𝗧𝗼𝗽 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀:

• 📁 𝗨𝗽𝗹𝗼𝗮𝗱 𝗹𝗮𝗿𝗴𝗲 𝗳𝗼𝗹𝗱𝗲𝗿𝘀 with ease using huggingface-cli upload-large-folder. Designed for your massive models and datasets. Much recommended if you struggle to upload your Llama 70B fine-tuned model 🤡
• 🔎 𝗦𝗲𝗮𝗿𝗰𝗵 𝗔𝗣𝗜: new search filters (gated status, inference status) and fetch trending score.
• ⚡𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲𝗖𝗹𝗶𝗲𝗻𝘁: major improvements simplifying chat completions and handling async tasks better.

We’ve also introduced tons of bug fixes and quality-of-life improvements - thanks to the awesome contributions from our community! 💪

💡 Check out the release notes: Wauplin/huggingface_hub#8

Want to try it out? Install the release with:

pip install huggingface_hub==0.25.0

1 reply

reacted to not-lain's post with 🔥 over 2 years ago

Post

2163

It is with great pleasure I inform you that huggingface's ModelHubMixin reached 200+ models on the hub 🥳

ModelHubMixin is a class developed by HF to integrate AI models with the hub with ease and it comes with 3 methods :
* save_pretrained
* from_pretrained
* push_to_hub

Shoutout to @nielsr , @Wauplin and everyone else on HF for their awesome work 🤗

If you are not familiar with ModelHubMixin and you are looking for extra resources you might consider :
* docs: https://huggingface.co/docs/huggingface_hub/main/en/package_reference/mixins
🔗blog about training models with the trainer API and using ModelHubMixin: https://huggingface.co/blog/not-lain/trainer-api-and-mixin-classes
🔗GitHub repo with pip integration: https://github.com/not-lain/PyTorchModelHubMixin-template
🔗basic guide: https://huggingface.co/posts/not-lain/884273241241808

reacted to clefourrier's post with 🔥 over 2 years ago

Post

6216

In a basic chatbots, errors are annoyances. In medical LLMs, errors can have life-threatening consequences 🩸

It's therefore vital to benchmark/follow advances in medical LLMs before even thinking about deployment.

This is why a small research team introduced a medical LLM leaderboard, to get reproducible and comparable results between LLMs, and allow everyone to follow advances in the field.

openlifescienceai/open_medical_llm_leaderboard

Congrats to @aaditya and @pminervini !
Learn more in the blog: https://huggingface.co/blog/leaderboard-medicalllm

Quentin Gallouédec PRO

AI & ML interests

Recent Activity

Organizations

qgallouedec's activity