Fine-tuning

A simple self-distillation method yields substantial code-generation gains across LLMs, boosting pass@1 on LiveCodeBench v6 from 42.4% to 55.3% and generalizing across Qwen and Llama models at 4B/8B/30B scales.

Show HN demonstrates fine-tuning Qwen2.5-7B on 100 films to build probabilistic story graphs.