ChristianAzinn/tiny-json

TinyJSON

Trained on my json-training dataset, these are finetunes of the smallest state-of-the-art LLMs to output in structured JSON.

Where their base/instruct versions have so little clue how to output JSON that forcing it using techniques like grammars simply hangs forever, these little guys (mostly) work like a charm. (SmolLM 135M still sometimes babbles on. Set a maximum token limit.)

Training was done with Unsloth at 4bit (lmao), rank=8, alpha=8, for 3 epochs each. rev1 models were trained on the first revision (11.6k rows) of json-training, while rev2 models were trained on the second (20.6k rows).