Liquid AI's 350M model gained 13.9 JSON points in 100 training steps

Hugging Face published a walkthrough on 3 September showing how a 350M-parameter model can be improved from 22.6 percent to 29.7 percent on IFStruct, a structured-output benchmark, using 100 steps of GRPO fine-tuning on about 500 samples. The gain on JSON specifically is 18.0 to 31.9 percent, or 13.9 points.

The setup

The base model is Liquid AI's LFM2.5-350M. The trainer is TRL's GRPO implementation with a LoRA adapter that touches 1.66 percent of the model's parameters. Training data is roughly 500 prompts from Nvidia's Nemotron-RL instruction-following structured-outputs dataset. Hugging Face reports the run fits on a single free-tier Colab or Kaggle GPU.

The reward functions

GRPO is a reinforcement-style method that scores each generation against reward functions. The post uses three, weighted 1.0, 0.5 and 2.0: whether the output parses as valid JSON, whether the field count matches the schema, and whether the payload passes JSON Schema validation. The heaviest weight is on schema conformance, which is where the biggest gain appears.

What the numbers say and do not say

The reported improvement is overall 22.6 to 29.7 on IFStruct, JSON 18.0 to 31.9, YAML 27.2 to 27.5, and bare list 16.6 to 29.7. Two things are worth reading carefully. First, the YAML score hardly changes. The reward functions were written for JSON, and the unchanged YAML score is evidence that the training targeted the right thing and left other formats alone. Second, 29.7 percent absolute is still a modest score. Read the result as a small model reaching a level of reliability where it can be trusted for narrow schema-conformance work. Performance of the leading models on general instruction following is a separate benchmark and this run does not measure it.

Hugging Face lists Leonie Monigatti as lead author, with contributors from Liquid AI and Hugging Face's own TRL team.

Teams that route requests to a large model only because a small local model kept returning malformed JSON now have a concrete recipe to test. A hundred training steps on 500 examples, running on free Colab, is a day of work for one engineer. If it works on the target schema, the saving applies to every later request: the small model runs on-device or on a cheap GPU, and the large model is no longer needed for the "produce a JSON object matching this shape" step.

The wider point is that GRPO with cheap, task-specific reward functions is now easy enough that a team without a machine-learning engineer can try it. The settings the post explains, from LoRA rank to reward weights to the KL penalty, are the ones that decide whether the trained model stays useful outside the benchmark, and they are worth tuning before drawing conclusions about your own workload.

Source

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