Post
2800
who's working on an NVFP4 version of Kimi-K3?
- 4 replies
AI & ML interests
None defined yet.
Recent Activity
View all activity
irenesolaimanย
authored a
paper 30 days ago
qgallouedecย
posted an update 3 months ago
Post
10616
Shipped hf-sandbox! ๐ฅก
๐งช 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:
๐ https://github.com/huggingface/hf-sandbox
๐งช 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:
๐ https://github.com/huggingface/hf-sandbox
- 1 reply
qgallouedecย
posted an update 3 months ago
Post
455
**TRL v1.4 is out ๐** Chunked NLL loss for SFT and a first-class **OpenReward** integration.
**Chunked NLL loss for SFT โ drops peak VRAM by up to 14ร**
Standard SFT materializes a full
Peak GPU memory, AdamW fp32:
- Qwen3-14B, 8รH100 FSDP2, 16k seq: 58.9 GB โ 38.9 GB
- Qwen3-4B, 1รH100 80GB, 16k seq: OOM โ 63.8 GB
- Qwen3-32B, 8รH100 FSDP2, 8k seq: OOM โ 71.2 GB
End-to-end it's consistently as fast or faster than
Works with PEFT and VLMs out of the box.
**Open Reward Standard environment adapter**
The new
v1.4 also brings MFU helpers for dense + MoE models, GRPO support for Liger 0.8.0 (delta clipping + VESPO + KL bias correction), Tรผlu 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
**Chunked NLL loss for SFT โ drops peak VRAM by up to 14ร**
Standard SFT materializes a full
[batch ร seq ร 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, 8รH100 FSDP2, 16k seq: 58.9 GB โ 38.9 GB
- Qwen3-4B, 1รH100 80GB, 16k seq: OOM โ 63.8 GB
- Qwen3-32B, 8รH100 FSDP2, 8k seq: OOM โ 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 โ a catalog name or a URL โ 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), Tรผlu 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
qgallouedecย
posted an update 3 months ago
Post
8168
TRL v1.3 ships day-one training support for Qwen 3.6 ๐
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 โ 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 โ 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
qgallouedecย
posted an update 3 months ago
Post
2077
TRL v1.2 introduces the SSDTrainer ๐
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.
v1.2 also ships expanded tool-calling support (LLaMA 3.1 / 3.2, DeepSeek-V3), another round of KTO โ DPO alignment getting us closer to promoting KTO to stable, a big GRPO simplification for overlong tool results, deprecation of
Full release notes: https://github.com/huggingface/trl/releases/tag/v1.2.0
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 โ 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
qgallouedecย
posted an update 4 months ago
Post
2491
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.
Read more: hf.co/blog/trl-v1
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 trlRead more: hf.co/blog/trl-v1
clefourrierย
authored a
paper 5 months ago
qgallouedecย
posted an update 5 months ago
Post
3106
@CohereLabs just released ๐ฟ Tiny Aya: a fully open-source 3B parameter model that speaks 70+ languages ๐! But thereโs a catch:
Tiny Aya is just a language model. It doesnโt 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 outโฆ itโs simple, thanks to Hugging Face TRL.
Weโre 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.
๐ https://github.com/huggingface/trl/blob/main/examples/notebooks/sft_tool_calling.ipynb
Tiny Aya is just a language model. It doesnโt 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 outโฆ itโs simple, thanks to Hugging Face TRL.
Weโre 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.
๐ https://github.com/huggingface/trl/blob/main/examples/notebooks/sft_tool_calling.ipynb
- 1 reply
lewtunย
submitted 2
papers to Daily Papers 6 months ago
julien-cย
submitted a
paper to Daily Papers 6 months ago
qgallouedecย
submitted a
paper to Daily Papers 7 months ago
nouamanetaziย
posted an update 9 months ago
Post
5057
After training ๐๐ฆ๐จ๐ฅ๐๐๐ on ๐๐๐ ๐๐๐๐๐ฌ for nearly a month, I've come to realize something most people overlook: ๐ข๐ง๐๐ซ๐๐ฌ๐ญ๐ซ๐ฎ๐๐ญ๐ฎ๐ซ๐ ๐ข๐ฌ ๐ญ๐ก๐ ๐ฆ๐๐ค๐-๐จ๐ซ-๐๐ซ๐๐๐ค ๐๐๐๐ญ๐จ๐ซ ๐ข๐ง ๐๐๐ ๐ญ๐ซ๐๐ข๐ง๐ข๐ง๐ . ๐ฅ
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious ๐๐๐๐ ๐๐ซ๐ซ๐จ๐ซ๐ฌ, or when your expensive GPU cluster is running at ๐๐% ๐๐๐๐ข๐๐ข๐๐ง๐๐ฒ, the problem isn't your model. It's most probably a ๐ฆ๐ข๐ฌ๐ฎ๐ฌ๐ ๐จ๐ ๐ญ๐ก๐ ๐ก๐๐ซ๐๐ฐ๐๐ซ๐. ๐ ๏ธ
Questions that seemed simple but had no clear answers: Why is ๐๐จ๐ ๐ญ๐ซ๐๐ข๐ง๐ข๐ง๐ ๐ฌ๐ฅ๐จ๐ฐ๐๐ซ ๐ญ๐ก๐๐ง ๐๐๐ง๐ฌ๐ ๐ฆ๐จ๐๐๐ฅ๐ฌ? Which ๐๐๐๐ ๐๐ฅ๐๐ ๐ฌ should we actually set? How often should we checkpoint without killing throughput?
That's why we built ๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค ๐: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the ๐ข๐ง๐๐ซ๐๐ฌ๐ญ๐ซ๐ฎ๐๐ญ๐ฎ๐ซ๐ ๐ฅ๐๐ฒ๐๐ซ that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: ๐๐๐๐ ๐ก๐ข๐ญ๐ญ๐ข๐ง๐ ๐ ๐๐/๐ฌ, ๐๐๐๐ข๐ง๐ค ๐.๐ ๐ซ๐๐๐๐ก๐ข๐ง๐ ๐๐๐ ๐๐/๐ฌ, ๐๐๐๐ ๐๐๐ง๐ ๐๐ญ ๐๐.๐ ๐๐/๐ฌ. Then we ran collective operations across ๐๐๐ ๐๐๐๐ฌ (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from ๐๐๐ ๐๐/๐ฌ on a single node to ๐๐๐-๐๐๐ ๐๐/๐ฌ across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค: https://lnkd.in/e5MKXUHS
Shared with โค๏ธ by the HuggingFace team
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious ๐๐๐๐ ๐๐ซ๐ซ๐จ๐ซ๐ฌ, or when your expensive GPU cluster is running at ๐๐% ๐๐๐๐ข๐๐ข๐๐ง๐๐ฒ, the problem isn't your model. It's most probably a ๐ฆ๐ข๐ฌ๐ฎ๐ฌ๐ ๐จ๐ ๐ญ๐ก๐ ๐ก๐๐ซ๐๐ฐ๐๐ซ๐. ๐ ๏ธ
Questions that seemed simple but had no clear answers: Why is ๐๐จ๐ ๐ญ๐ซ๐๐ข๐ง๐ข๐ง๐ ๐ฌ๐ฅ๐จ๐ฐ๐๐ซ ๐ญ๐ก๐๐ง ๐๐๐ง๐ฌ๐ ๐ฆ๐จ๐๐๐ฅ๐ฌ? Which ๐๐๐๐ ๐๐ฅ๐๐ ๐ฌ should we actually set? How often should we checkpoint without killing throughput?
That's why we built ๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค ๐: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the ๐ข๐ง๐๐ซ๐๐ฌ๐ญ๐ซ๐ฎ๐๐ญ๐ฎ๐ซ๐ ๐ฅ๐๐ฒ๐๐ซ that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: ๐๐๐๐ ๐ก๐ข๐ญ๐ญ๐ข๐ง๐ ๐ ๐๐/๐ฌ, ๐๐๐๐ข๐ง๐ค ๐.๐ ๐ซ๐๐๐๐ก๐ข๐ง๐ ๐๐๐ ๐๐/๐ฌ, ๐๐๐๐ ๐๐๐ง๐ ๐๐ญ ๐๐.๐ ๐๐/๐ฌ. Then we ran collective operations across ๐๐๐ ๐๐๐๐ฌ (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from ๐๐๐ ๐๐/๐ฌ on a single node to ๐๐๐-๐๐๐ ๐๐/๐ฌ across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค: https://lnkd.in/e5MKXUHS
Shared with โค๏ธ by the HuggingFace team
Post
4443
๐ค Did you know your voice might be cloned without your consent from just *one sentence* of audio?
That's not great. So with @frimelle , we brainstormed a new idea for developers who want to curb malicious use: โจThe Voice Consent Gate.โจ
Details, code, here: https://huggingface.co/blog/voice-consent-gate
That's not great. So with @frimelle , we brainstormed a new idea for developers who want to curb malicious use: โจThe Voice Consent Gate.โจ
Details, code, here: https://huggingface.co/blog/voice-consent-gate
- 3 replies
anditoย
authored a
paper 9 months ago
Post
3163
Finally, our new paper is out! "๐๐ถ๐ป๐ฒ๐ฉ๐ถ๐๐ถ๐ผ๐ป: ๐ข๐ฝ๐ฒ๐ป ๐๐ฎ๐๐ฎ ๐๐ ๐๐น๐น ๐ฌ๐ผ๐ ๐ก๐ฒ๐ฒ๐ฑ"! ๐ฅณ
FineVision: Open Data Is All You Need (2510.17269)
If you've ever trained a VLM, you know this problem: nobody shares their data mixtures. It's a black box, making replicating SOTA work impossible.
We wanted to change that.
FineVision unifies 200 sources into 24 million samples. With 17.3 million images and 9.5 billion answer tokens, it's the largest open resource of its kind.
In the paper, we share how we built it:
๐ finding and cleaning data at scale
๐งน removing excessive duplicates across sources
๐ค decontaminating against 66 public benchmarks
My favorite part is Figure 6 (in the video!). It's our visual diversity analysis. It shows that FineVision isn't just bigger; it's more balanced and conceptually richer than other open datasets.
NVIDIA's Eagle 2 paper highlighted just how critical this visual diversity is, and our results confirm it: models trained on FineVision consistently outperform those trained on any other open dataset on 11 benchmarks!
๐ To celebrate the paper, Iโm also releasing a concatenated and shuffled version of the full dataset! ๐
Itโs ready to stream, so you can start training your own models right away:
from datasets import load_dataset
d = load_dataset("HuggingFaceM4/FineVision_full_shuffled", split="train", streaming=True)
print(next(iter(d)))
A big shoutout to the first authors: Luis Wiedmann and Orr Zohar. They are rockstars!
FineVision: Open Data Is All You Need (2510.17269)
If you've ever trained a VLM, you know this problem: nobody shares their data mixtures. It's a black box, making replicating SOTA work impossible.
We wanted to change that.
FineVision unifies 200 sources into 24 million samples. With 17.3 million images and 9.5 billion answer tokens, it's the largest open resource of its kind.
In the paper, we share how we built it:
๐ finding and cleaning data at scale
๐งน removing excessive duplicates across sources
๐ค decontaminating against 66 public benchmarks
My favorite part is Figure 6 (in the video!). It's our visual diversity analysis. It shows that FineVision isn't just bigger; it's more balanced and conceptually richer than other open datasets.
NVIDIA's Eagle 2 paper highlighted just how critical this visual diversity is, and our results confirm it: models trained on FineVision consistently outperform those trained on any other open dataset on 11 benchmarks!
๐ To celebrate the paper, Iโm also releasing a concatenated and shuffled version of the full dataset! ๐
HuggingFaceM4/FineVision_full_shuffled Itโs ready to stream, so you can start training your own models right away:
from datasets import load_dataset
d = load_dataset("HuggingFaceM4/FineVision_full_shuffled", split="train", streaming=True)
print(next(iter(d)))
A big shoutout to the first authors: Luis Wiedmann and Orr Zohar. They are rockstars!
- 4 replies
ariG23498ย
authored a
paper 9 months ago
thomwolfย
authored a
paper 10 months ago
lvwerraย
authored a
paper 10 months ago