Qwen3.5-9B — Blog-Provider-ID — Cold-Start RL (SFT-60 → reasoning GRPO)
Method: generic reasoning GRPO (control DPPO + Dr.GRPO, hint-free prompt, no teacher/cheatsheet) initialised from the 60-step cold-start SFT checkpoint. The headline of the cold-start ablation.
Result: the first reasoning-channel RL here that climbs AND never collapses — val ~0.67 → ~0.89 with 0% truncation at every step (base-init reasoning RL collapses to 60–100% truncation every seed). Shows the collapse was a cold-start pathology, not intrinsic to the reasoning channel. Still below answer-only SFT (1.000); value is diagnostic.
- 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
