A guide demonstrates fine-tuning the Liquid AI LFM2.5-350M model using Group Relative Policy Optimization (GRPO) to enhance structured output generation. The process involves training on a subset of the NVIDIA Nemotron dataset and evaluating performance via the IFStruct benchmark.
- Baseline evaluation of the base LFM2.5-350M model yielded a 22.6% pass rate on the IFStruct benchmark.
- Training utilized approximately 500 samples with data augmentation to align with the evaluation distribution.
- Reward functions were combined to score JSON format, field count accuracy, and schema validation.
- The fine-tuned model achieved a 29.7% overall pass rate, representing a significant improvement over the baseline.
The results show that task-specific fine-tuning of smaller models can improve performance and match that of far larger models.