Qwen3-4B-Instruct-2507 distilled on gpt-5.6-luna thoughts about the next 8 Qwen3 tokens of stat.ML arXiv LaTeX.

SFT: 1 epoch over 604,766 examples, lr 1e-5 cosine, batch 256, bf16. Final val log prob on held-out luna thoughts: -0.9238 nats/token (from -3.31 at step 1).

Prompt it with the SFT chat template shipped here as chat_template.jinja — luna's own v5trim-qwen system prompt minus its few-shot exchanges. Generation stops on <|im_end|>. Recommended sampling (Qwen3-2507): temperature 0.7, top_p 0.8, top_k 20, min_p 0.

Thoughts: JackHsieh/luna-reason-only.k-8.statml-arxiv-qwen3