Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps

There is something almost thermodynamic about this result. A language model, left to its own devices, is a seething bath of tokens — an entropy machine that predicts the next word from a blur of prior text. But throw in GRPO — Group Relative Policy Optimization — and suddenly that same model crystallizes into a rigorous formatter, emitting valid JSON with the inevitability of a falling apple. It's as if the second law of thermodynamics got a loophole: local order, purchased cheaply, by comparing a handful of sampled futures against each other. What makes GRPO so philosophically delicious is its refusal of absolute judgment. There's no external critic, no oracle whispering the "correct" answer. Instead, the model samples a group of possible responses, ranks them relative to one another, and nudges itself toward the better ones. This is natural selection at the level of thought — a Darwinian tournament where mediocre outputs are the fossils, and well-structured ones get to reproduce their gradients. In a universe that runs on relative comparisons — entropy gradients, potential differences, fitness landscapes — this feels less like a hack and more like a fundamental law of learning. The efficiency is the mind-bender. One hundred steps. A 350M model, which is practically a pebble next to the boulders of modern LLMs, learns structured output behavior that typically requires either massive scale or painstaking supervised fine-tuning. This hints at something deep about how neural networks represent rules: the latent space already contains the skeleton of grammar, logic, and schema — it just needs the right pressure to snap into place. Like a supercooled liquid that freezes the instant it's disturbed, the model was always one good reward signal away from rigidity. For the Weird & Wild reader, the takeaway is cosmic: structure is not imposed from above, it emerges from competition. Whether it's quarks binding into protons, cells organizing into brains, or tokens aligning into JSON, the universe seems to prefer patterns that are selected, not designed. This blog post from Hugging Face — — is a practical recipe, but it's also a parable. If a 350M model can learn to