A 100-step reinforcement learning run raised the structured-output accuracy of Liquid AI's 350-million-parameter model from 22.6% to 29.7%, Hugging Face reported on Sept. 3, 2026.
Researchers fine-tuned the LFM2.5-350M model using Group Relative Policy Optimization through the TRL library. The workflow trained roughly 6 million parameters—about 1.66% of the model—by applying LoRA adapters across the hybrid attention and convolution architecture.
Training relied on about 500 samples from the nvidia/Nemotron-RL-instruction_following-structured_outputs dataset. Hugging Face appended fenced code block instructions to 40% of the prompts and converted a disjoint 20% into top-level array tasks. Three weighted rewards scored completions on format compliance, top-level field counts, and JSON Schema validation.
Evaluation took place on the 2,000-sample IFStruct benchmark using llama.cpp on an Apple M5 Max MacBook Pro. In the baseline run, the base model passed 452 tasks, scoring 18.0% on JSON and 27.2% on YAML. After 100 steps of GRPO, the fine-tuned checkpoint passed 594 tasks, lifting the JSON pass rate to 31.9% while YAML accuracy reached 27.5%.
The fine-tuned model recorded an average latency of 1,518 milliseconds, compared to 1,453 milliseconds for the base checkpoint. The final 29.7% score remains below Qwen3.5-2B's 33.15% result on IFStruct, though the training process ran on a single 16-gigabyte GPU.
