open-acc (open/ acc)

posted an update 7 days ago

Post

428

did you know you can train agentic models with RL deploying the environments on HF Spaces? 🤗

with TRL + OpenEnv, your training script connects to remote environments hosted as Spaces

want to train faster? → just add more Spaces (TRL handles the parallelization natively)

we used this to train a model to solve the trolley problem in CARLA. 2 HF Spaces running a full driving simulator, each on a T4 GPU

full write-up with code and results → https://huggingface.co/blog/sergiopaniego/bringing-carla-to-openenv-trl

sergiopaniego

posted an update 8 days ago

Post

328

Qwen3.5 dense (smol 🤏) models just dropped

- natively multimodal
- 0.8B · 2B · 4B · 9B (+ base variants)
- 262K context extensible to 1M
- built-in thinking

fine-tune them with TRL out of the box → SFT, GRPO, DPO and more!

examples: https://huggingface.co/docs/trl/example_overview
collection: https://huggingface.co/collections/Qwen/qwen35

sergiopaniego

posted an update 12 days ago

Post

2309

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

mitkox

posted an update 17 days ago

Post

5348

My USB charger has a Blackwell GPU and 128GB RAM.
What. A. Time. To. Be. Alive.
People in Sofia: “It’s freezing.”
Me: sitting next to 3kW of space AI heaters on my desk 👀
1x GLM-5, 2x MiniMax-M2.5, 1x Qwen3 Coder Next; all on single Aibrix/K8s cluster

6 replies

·

sergiopaniego

posted an update 20 days ago

Post

1445

Tiny Aya 🌿 just dropped from @CohereLabs , a really powerful multilingual small model!

To celebrate, we cooked up fresh resources to train it for tool calling 🔧

> Free Google Colab guide: https://colab.research.google.com/github/huggingface/trl/blob/main/examples/notebooks/sft_tool_calling.ipynb
> Standalone training script: https://github.com/huggingface/trl/blob/main/examples/scripts/sft_tiny_aya_tool_calling.py

mitkox

posted an update 20 days ago

Post

291

134,614 tok/sec input prefil max
1031 tokens/sec out gen max

At these local AI speeds, there is no User Interface for humans. My human UI is the Radicle distributed Git issues queue

On my GPU workstation:
- Z8 Fury G5 4x A6000
- MiniMax-M2.5
- Claude Code to localhost:8000

1 reply

·

sergiopaniego

posted an update 25 days ago

Post

544

The latest piece by @MiniMax-AI is a must-read.

It tries to break the impossible triangle of agent RL: throughput × stability × flexibility.

A lot to learn here, go read it 🫵
https://huggingface.co/blog/MiniMax-AI/forge-scalable-agent-rl-framework-and-algorithm

1aurent

authored a paper 27 days ago

Ministral 3

Paper • 2601.08584 • Published Jan 13 • 55

sergiopaniego

posted an update 29 days ago

Post

493

if you're looking for a good first issue to get your open-source journey started, you could contribute to this TRL issue by documenting one impactful paper in the docs

we have a broad list to cover!! 🧐

https://github.com/huggingface/trl/issues/4407

mitkox

posted an update about 1 month ago

Post

4734

I just pushed Claude Code Agent Swarm with 20 coding agents on my desktop GPU workstation.

With local AI, I don’t have /fast CC switch, but I have /absurdlyfast:
- 100’499 tokens/second read, yeah 100k, not a typo | 811 tok/sec generation
- KV cache: 707’200 tokens
- Hardware: 5+ year old GPUs 4xA6K gen1; It’s not the car. It’s the driver.

Qwen3 Coder Next AWQ with cache at BF16. Scores 82.1% in C# on 29-years-in-dev codebase vs Opus 4.5 at only 57.5%. When your codebase predates Stack Overflow, you don't need the biggest model; you need the one that actually remembers Windows 95.

My current bottleneck is my 27" monitor. Can't fit all 20 Theos on screen without squinting.

3 replies

·

Sri-Vigneshwar-DJ

posted an update about 1 month ago

Post

1402

Just released a new dataset designed for training reasoning models on Meta (Facebook/Instagram) advertising fatigue detection!

What is it? A GRPO (Group Relative Policy Optimization) training dataset with 200+ carefully crafted scenarios covering:

🔍 Fatigue Signal Detection: CTR drops, CPM spikes, frequency analysis
🩺 Performance Diagnosis: Root cause analysis frameworks
📋 Strategy: Creative refresh cadence, testing frameworks
📊 Analysis: ROI calculations, metric interpretation
Why GRPO? GRPO training helps models learn structured reasoning. Each response follows the <thinking> and <answer> format.

Check it out here: Sri-Vigneshwar-DJ/meta-fatigue-grpo-dataset

julien-c

submitted a paper to Daily Papers about 1 month ago

Shaping capabilities with token-level data filtering

Paper • 2601.21571 • Published Jan 29 • 27

sergiopaniego

posted an update about 1 month ago

Post

531

Meet the Post-Training Toolkit (PTT), which easily integrates with TRL via a single callback, by Aditya Challapally ( @microsoft ):

🔍 Detects training issues early
🛠 Lets you intervene safely
📊 Keeps long training runs stable, auditable & efficient

Microsoft blog: https://devblogs.microsoft.com/engineering-at-microsoft/diagnosing-instability-in-production-scale-agent-rl/

Integration guide: https://huggingface.co/docs/trl/main/en/ptt_integration

Code: https://github.com/microsoft/post-training-toolkit

mitkox

posted an update about 1 month ago

Post

312

▐▛██▜▌ Claude Code v2.1.23
▝████▘ Kimi-K2.5 · API Usage Billing
▘▘ ▝▝ ~/dev/vllm
/model to try Opus 4.5
❯ hey
● Hello! How can I help you today?
❯ what model are you?
● I'm Claude Kimi-K2.5, running in a local environment on Linux.

Took some time to download and vLLM hybrid inferencing magic to get it running on my desktop workstation.

sergiopaniego

posted an update about 1 month ago

Post

2594

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

·

sergiopaniego

posted an update about 1 month ago

Post

2205

Date idea: read the entire Transformers v5.0.0 release notes

Officially stable now: https://github.com/huggingface/transformers/releases/tag/v5.0.0

1 reply

·

Sri-Vigneshwar-DJ

posted an update about 1 month ago

Post

219

🏙️ Hugging Face Community Post
Title: 🧬 Experimenting with "Dynamic Chaos" in Tamil SLMs

Hi everyone! I just published a new experimental study on Small Language Model (SLM) resilience.

I took the Qwen2.5-0.5B model and put it through a "Chaos Phase" to see how much weight data a tiny model can lose before its understanding of classical Tamil grammar breaks.

Key highlights of the study:

Target Data: Fine-tuned on the Thirukkural (1,330 couplets + modern explanations).
The Chaos Step: Applied 20% random weight pruning but implemented "Layer Protection" for the Token Embeddings and LM Head to keep the characters readable.
Compression: 4-bit (Q4_K_M) quantization for extreme efficiency.
Result: A surrealist classical Tamil model that is ultra-light (~300MB) and ultra-fast!

Check out the model and the experiment logic here: Sri-Vigneshwar-DJ/qwen-tamil-chaos-v1

mitkox

posted an update about 2 months ago

Post

1539

GLM-4.7-Flash is fast, good and cheap.
3,074 tokens/sec peak at 200k tokens context window on my desktop PC.
Works with Claude Code and opencode for hours. No errors, drop-in replacement of the Anthropic cloud AI.
MIT licensed, open weights, free for commercial use and modifications.
Supports speculative decoding using MTP, which is highly effective in mitigating latency.
Great for on device AI coding as AWQ 4bit at 18.5 GB. Hybrid inference on a single consumer GPU + CPU RAM.

3 replies

·

MElHuseyni

authored 2 papers about 2 months ago

Beyond Cosine Similarity: Taming Semantic Drift and Antonym Intrusion in a 15-Million Node Turkish Synonym Graph

Paper • 2601.13251 • Published Jan 19 • 4

A Hybrid Protocol for Large-Scale Semantic Dataset Generation in Low-Resource Languages: The Turkish Semantic Relations Corpus

Paper • 2601.13253 • Published Jan 19 • 4

open/ acc

AI & ML interests

Recent Activity

Ministral 3

Shaping capabilities with token-level data filtering

Beyond Cosine Similarity: Taming Semantic Drift and Antonym Intrusion in a 15-Million Node Turkish Synonym Graph

A Hybrid Protocol for Large-Scale Semantic Dataset Generation in Low-Resource Languages: The Turkish Semantic Relations Corpus

AI & ML interests

Recent Activity

Team members 217

open-acc's activity