# Hugging Face — Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps

- Company: Hugging Face (huggingface.co)
- Announced: 2026-09-03
- Category: developer-tool-release
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://huggingface.co/blog/grpo-with-trl-ifstruct
- Record: https://forck.live/items/7607-fine-tuning-a-350m-model-for-better-structured-outputs-in-100-grpo-steps
- Subject: Platform
- Open weights: yes
- Models affected: LFM2.5-350M
- Context window: 32768

Hugging Face published an open-source recipe and notebook to fine-tune Liquid AI's LFM2.5-350M model using GRPO via the TRL library, improving structured-output compliance from 22.6% to 29.7% on the IFStruct benchmark. The guide runs on a free-tier Colab or Kaggle GPU using 500 samples and 100 training steps. Materials are available on GitHub.

## Evidence

Verbatim from https://huggingface.co/blog/grpo-with-trl-ifstruct:

> The full run takes around 500 samples and 100 training steps, small enough for a free-tier Colab or Kaggle GPU, and is available on GitHub.

---

Record: https://forck.live/items/7607-fine-tuning-a-350m-model-for-better-structured-outputs-in-100-grpo-steps
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md