Agentic RL: Token-In, Token-Out Done Right
• 11
AI & ML interests
The AI community building the future.
Recent Activity
View all activity
Papers
Seeing the Needle in the Haystack: Towards Weakly-Supervised Log Instance Anomaly Localization via Counterfactual Perturbation
Qualixar OS: A Universal Operating System for AI Agent Orchestration
Articles
jeffboudier
updated a
Space about 2 hours ago
abidlabs
updated a
Space about 3 hours ago
Images for docs PR 3300
#625 opened about 4 hours ago
by 
hubnemo
lysandre
updated a
dataset about 12 hours ago
evalstate
updated a
bucket about 13 hours ago
sayakpaul
updated a
dataset about 21 hours ago
alvarobartt
updated a
dataset about 22 hours ago
ariG23498
updated a
dataset 1 day ago
Post
1007
Weekend mini project! Since commentary on AI is inherently interdisciplinary, we connected the observations in the Pope's encyclical with decades of scholarship in Responsible AI and Ethics research and created an interactive space with these annotations!
Work with @IJ-Reynolds , @yjernite , and @meg
Lots to unpack. We started with 105 annotations. Please submit pull requests for more that we may have missed!
society-ethics/annotated-encyclical
Work with @IJ-Reynolds , @yjernite , and @meg
Lots to unpack. We started with 105 annotations. Please submit pull requests for more that we may have missed!
society-ethics/annotated-encyclical
alvarobartt
posted an update 15 days ago
Post
316
Open agents on AWS SageMaker AI with open models from the Hugging Face Hub!
> Deploy an open model from the Hugging Face Hub on SageMaker AI
> Connect the deployed model to Strands Agents
> Add built-in and custom tools for tool calling
> Expose external capabilities through MCP integration
> Bonus: talk to your agent and visualize traces with Gradio
https://alvarobartt.com/agents-on-aws-sagemaker
> Deploy an open model from the Hugging Face Hub on SageMaker AI
> Connect the deployed model to Strands Agents
> Add built-in and custom tools for tool calling
> Expose external capabilities through MCP integration
> Bonus: talk to your agent and visualize traces with Gradio
https://alvarobartt.com/agents-on-aws-sagemaker
Post
198
Two large changes in kernel-builder this week:
kernel-builder now links libstdc++ dynamically. To support a wide range of systems, we build against libstdc++ from manylinux_2_28 (EL 8 and later).
Following our Torch support policy that the current and previous Torch versions are supported, Torch 2.10 support was removed. We will soon also support the Torch stable ABI, so that it is possible to write kernels that support a large number of Torch versions.
kernel-builder now links libstdc++ dynamically. To support a wide range of systems, we build against libstdc++ from manylinux_2_28 (EL 8 and later).
Following our Torch support policy that the current and previous Torch versions are supported, Torch 2.10 support was removed. We will soon also support the Torch stable ABI, so that it is possible to write kernels that support a large number of Torch versions.
Post
706
🤗 Announcing the Ettin Reranker family: six new state-of-the-art CrossEncoder rerankers for search from 17M to 1B parameters, plus the full training data and the ~150-line recipe. Built on the Ettin ModernBERT encoders, Apache 2.0. Details:
All six were trained with the same single-stage pointwise MSE distillation recipe, with mixedbread-ai/mxbai-rerank-large-v2 (1.54B) as the teacher. Only the learning rate and per-device batch size change between sizes. The 1B student matches the teacher within 0.0001 NDCG@10 on MTEB(eng, v2) Retrieval, the 150M is the strongest reranker I tested in the under-600M range, and the 17M beats the 33M ms-marco-MiniLM-L12-v2 by +0.051 NDCG@10 at roughly half the parameter count.
Speed matters as much as quality for a reranker, since it determines whether the model fits the latency budget between retrieval and showing results. Our 17M is the fastest reranker in the whole comparison at 7517 pairs/sec on an H100. Our 150M runs 2.3x faster than the two other 150M ModernBERT-base rerankers (gte-reranker-modernbert-base and granite-embedding-reranker-english-r2) because the modular Transformer module propagates unpadded inputs through every layer rather than just the FA2 attention kernel. And our 1B is 2.4x faster than its 1.5B teacher while matching it on quality.
I bootstrapped the training recipe with the new train-sentence-transformers Agent Skill shipped in Sentence Transformers v5.5.0. Install it with
I wrote a blog post walking through usage, results across six embedder pairings, the speed story, and the complete training script. Check it out, or just point your Agent to the URL:
https://huggingface.co/blog/ettin-reranker
Collection: https://huggingface.co/collections/cross-encoder/ettin-rerankers
All six were trained with the same single-stage pointwise MSE distillation recipe, with mixedbread-ai/mxbai-rerank-large-v2 (1.54B) as the teacher. Only the learning rate and per-device batch size change between sizes. The 1B student matches the teacher within 0.0001 NDCG@10 on MTEB(eng, v2) Retrieval, the 150M is the strongest reranker I tested in the under-600M range, and the 17M beats the 33M ms-marco-MiniLM-L12-v2 by +0.051 NDCG@10 at roughly half the parameter count.
Speed matters as much as quality for a reranker, since it determines whether the model fits the latency budget between retrieval and showing results. Our 17M is the fastest reranker in the whole comparison at 7517 pairs/sec on an H100. Our 150M runs 2.3x faster than the two other 150M ModernBERT-base rerankers (gte-reranker-modernbert-base and granite-embedding-reranker-english-r2) because the modular Transformer module propagates unpadded inputs through every layer rather than just the FA2 attention kernel. And our 1B is 2.4x faster than its 1.5B teacher while matching it on quality.
I bootstrapped the training recipe with the new train-sentence-transformers Agent Skill shipped in Sentence Transformers v5.5.0. Install it with
hf skills add train-sentence-transformers --claude and ask Claude Code (or Codex / Cursor / Gemini CLI) to fine-tune a SentenceTransformer, CrossEncoder, or SparseEncoder model on your data.I wrote a blog post walking through usage, results across six embedder pairings, the speed story, and the complete training script. Check it out, or just point your Agent to the URL:
https://huggingface.co/blog/ettin-reranker
Collection: https://huggingface.co/collections/cross-encoder/ettin-rerankers
alvarobartt
posted an update 18 days ago
Post
3289
Latest
TL; DR MoEs can be misleading to reason about from active parameters alone, since each token only activates a subset of experts, while the serving setup still needs to account for the full resident memory footprint.
🧠
🏗️ Dense models usually load and use most weights every forward pass, while MoEs load many experts but only route each token to a few of them
⚡ Active params isn't the same as memory footprint, especially for sparse architectures
📦 Runtime memory is about what is used per request/token, while loading memory also includes the expert weights that need to be resident
📚 KV cache can still dominate depending on context length, batch size, and concurrency
🔀 Expert Parallelism (EP) helps shard experts across accelerators when expert weights dominate
🚀 Data Parallelism (DP) + EP is often a good fit for throughput-oriented MoE serving
Check the repository at https://github.com/alvarobartt/hf-mem
hf-mem release added a breakdown of Mixture-of-Experts (MoE) memory usage!TL; DR MoEs can be misleading to reason about from active parameters alone, since each token only activates a subset of experts, while the serving setup still needs to account for the full resident memory footprint.
🧠
hf-mem now splits MoE memory into base model weights, routed experts, and KV cache🏗️ Dense models usually load and use most weights every forward pass, while MoEs load many experts but only route each token to a few of them
⚡ Active params isn't the same as memory footprint, especially for sparse architectures
📦 Runtime memory is about what is used per request/token, while loading memory also includes the expert weights that need to be resident
📚 KV cache can still dominate depending on context length, batch size, and concurrency
🔀 Expert Parallelism (EP) helps shard experts across accelerators when expert weights dominate
🚀 Data Parallelism (DP) + EP is often a good fit for throughput-oriented MoE serving
Check the repository at https://github.com/alvarobartt/hf-mem
Post
440
🤖 I've just published Sentence Transformers v5.5.0, headlined by a new train-sentence-transformers Agent Skill that lets your AI coding agent (Claude Code, Codex, Cursor, Gemini CLI, ...) train and finetune embedding, reranker, and sparse encoder models for you. Plus training losses & fixes. Details:
The skill bundles curated guidance for the whole training workflow across all three model types: base model selection, loss and evaluator choice, hard-negative mining, distillation, LoRA, Matryoshka, multilingual training, static embeddings, etc. It also ships production-ready training template scripts the agent can adapt. Install it with
On the loss side: EmbedDistillLoss is a new embedding-level distillation loss for SentenceTransformer. Instead of distilling teacher scores like MarginMSELoss, it aligns the student's embeddings directly with pre-computed teacher embeddings, wtih an optional learnable projection for when the student and teacher dimensions differ. Second, ADRMSELoss is a new listwise learning-to-rank loss for CrossEncoder from the Rank-DistilLLM paper, aimed at the LLM-distillation reranking setting.
encode() and predict() also gained a per-call processing_kwargs override, so you can change processor settings like max_length, a vision-language model's image resolution, or a video's fps, for a single call without rebuilding the model.
The Agent Skill is the part of this release I'm most keen for people to try. Curious to hear how it works for you. I've been using it myself a lot to quickly set up some training runs that immediately use a bunch of best practices.
> pip install sentence-transformers==5.5.0
> hf skills add train-sentence-transformers
The full release notes: https://github.com/huggingface/sentence-transformers/releases/tag/v5.5.0
The skill bundles curated guidance for the whole training workflow across all three model types: base model selection, loss and evaluator choice, hard-negative mining, distillation, LoRA, Matryoshka, multilingual training, static embeddings, etc. It also ships production-ready training template scripts the agent can adapt. Install it with
hf skills add train-sentence-transformers, then just describe what you want, e.g. "finetune a reranker on my (question, answer) pairs, mine hard negatives, and push it to the Hub".On the loss side: EmbedDistillLoss is a new embedding-level distillation loss for SentenceTransformer. Instead of distilling teacher scores like MarginMSELoss, it aligns the student's embeddings directly with pre-computed teacher embeddings, wtih an optional learnable projection for when the student and teacher dimensions differ. Second, ADRMSELoss is a new listwise learning-to-rank loss for CrossEncoder from the Rank-DistilLLM paper, aimed at the LLM-distillation reranking setting.
encode() and predict() also gained a per-call processing_kwargs override, so you can change processor settings like max_length, a vision-language model's image resolution, or a video's fps, for a single call without rebuilding the model.
The Agent Skill is the part of this release I'm most keen for people to try. Curious to hear how it works for you. I've been using it myself a lot to quickly set up some training runs that immediately use a bunch of best practices.
> pip install sentence-transformers==5.5.0
> hf skills add train-sentence-transformers
The full release notes: https://github.com/huggingface/sentence-transformers/releases/tag/v5.5.0
- 3 replies
evijit
submitted a
paper to Daily Papers 25 days ago
qgallouedec
posted an update 26 days ago
Post
10230
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
🧪 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
qgallouedec
posted an update 28 days ago
Post
371
**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
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
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), 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
**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
qgallouedec
posted an update about 1 month ago
Post
8085
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
AdinaY
submitted a
paper to Daily Papers about 2 months ago
qgallouedec
posted an update about 2 months ago
Post
2031
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.
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
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 ↔ 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