RL-Repaired-V1

Six inference/evaluation LoRA adapters at optimizer step 125: Qwen3.5 M0-v4 4B and 9B crossed with DPO, GRPO, and PPO. Every run used the exact ordered prompts-repaired-v1 cohort and its versioned localverifier-v1 derivative.

Size	DPO	GRPO	PPO
4B	4b/dpo/step-125	4b/grpo/step-125	4b/ppo/step-125
9B	9b/dpo/step-125	9b/grpo/step-125	9b/ppo/step-125

The adapter configs point to immutable M0-v4 bases. Optimizer states are not present in the source checkpoint directories and are therefore not included; these are loadable model-only checkpoints, not training-resume archives. See each artifact_manifest.json and the root upload_manifest.json for hashes and provenance.