Qwen3.5-9B — Blog-Provider-ID — RLSD (E3, verifier-anchored self-distillation)
Method (OPSD E3): RLSD — on-policy self-distillation that combines the verifier sign (direction) with a cheatsheet-teacher per-token magnitude (coef_t = sign(A)·exp(clip(sign(A)·Δ))), PPO-clipped on the student ratio. Teacher = the live student pool with the train-derived style cheatsheet spliced into the scoring prompt. Init from base. Final checkpoint (step_40).
Result: produced the best transient reasoning-channel number (peaked val ~0.41, truncation briefly 38%→9%) then collapsed into ~100% truncation without a trust-region/length anchor. Part of the reasoning-channel-RL-collapse analysis.
- 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
