Sakhanda Wire
NVDA $153.59 -1.64% MSFT $507.00 -2.12% GOOGL $279.98 -1.32% META $619.94 -1.03% AMZN $256.62 -0.44%
← Back to the news

GRPO fine-tuning lifts LFM2.5-350M structured-output score

A public training recipe uses Group Relative Policy Optimization (GRPO) to improve structured-output compliance in LiquidAI’s LFM2.5-350M model. The procedure relies on the TRL library, about 500 training samples and 100 steps, allowing the run to fit a free-tier Colab or Kaggle GPU. Evaluation uses the open IFStruct benchmark, which checks whether responses are valid and conform to requested schemas and structures.

In the reported local setup, the base model passed 452 of 2,000 tests, or 22.6%. After fine-tuning with a LoRA adapter, the merged model passed 594 tests, reaching 29.7%, a gain of 7.1 percentage points. The JSON pass rate increased from 18.0% to 31.9%, while YAML changed from 27.2% to 27.5%. Bare-list compliance rose from 16.6% to 29.7%.

The training data paired prompts with JSON schemas and expected field counts. Some examples were adapted to require fenced code blocks or top-level arrays. The authors note that this pipeline is separate from the one used for the reinforcement-learning model described in the IFStruct blog and is intended to demonstrate task-specific gains in a smaller model.

Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.

Originally published by Hugging Face on

Read the original on Hugging Face ↗

The reporting and the image belong to Hugging Face and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news