CK0607/qwen3.5-9b-blogprovider-selfgated-answeronly

Qwen3.5-9B — Blog-Provider-ID — Self-Gated Answer-Only SFT

Method (E1 control): answer-only SFT on 1071 verifier-gated rows (the model's own rollouts; gold only verifier-gates which are kept, never shown during generation; rationale STRIPPED, only the label supervised).

Result: val 1.000 / val_ood 1.000. Headline finding: answer-only BEATS the with-reasoning STaR run (0.953) on identical data and MATCHES the gold-conditioned SFT (1.000) with fewer rows and no rationale — on this task the discriminative signal is a dense supervised label-mapping, and supervising self-generated rationales is net-harmful.

• 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