Qwen3.5-9B — Blog-Provider-ID — RL (pure-accuracy, peak)
Method: answer-only / no-reasoning GRPO (reward = exact label match). This is the PEAK checkpoint (step_16) before the run collapsed.
Result: peaked val 0.650 / val_ood 0.662, then collapsed (CLAUDE absorbed into CHATGPT → 2-class ceiling) and hit the zero-trainable-batch guardrail. Directionally showed answer-only > reasoning-RL, but sparse single-token RL collapsed the fine boundary. (NB: raw weight-broadcast snapshot; config/tokenizer copied from the matching base arch.)
- 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
