CK0607/qwen3.5-9b-blogprovider-coldstart-sft

Qwen3.5-9B — Blog-Provider-ID — Cold-Start SFT warm-up (60 steps)

Method: 60-step SFT from base on 1071 verifier-passed STaR reasoning traces (<reason_why>/<answer>). The Phase-1 warm-start for the cold-start RL ablation — teaches the reasoning format + a competent prior, not to solve the task.

Result: lifts the reasoning-channel starting policy to val ~0.67 at 0% truncation (vs ~0.13–0.21 for base init). Used to initialise the cold-start RL run (see -coldstart-rl).

• Base model: Qwen/Qwen3.5-9B (thinking OFF)
• Task: 3-way AI-provider classification — given a blog/essay, identify whether it was written by CLAUDE, CHATGPT, or GEMINI. Output format: <reason_why>...</reason_why><answer>LABEL\nConfidence: ...</answer>.
• Eval: val (in-distribution topics, n=414) and val_ood (held-out topics, n=471), zero eval leakage.
• Provenance: prime-rl; code at https://github.com/ChinmayK0607/prime-rl/tree/blog-author-id-experiments