qwen3-8b_chem_rlcr
Research checkpoints for the RLCR readout project. Each iter_* subfolder is a complete
Hugging Face checkpoint at the indicated completed rollout count. RLVR and RLCR use the
same answer-and-confidence prompt and CISPO optimization. These are experimental weights;
mechanistic conclusions require the project's corrected evaluations and controls.
Each folder includes validation.json with CPU tensor checks and SHA-256 file hashes.
For Qwen3.5, visual and disabled MTP tensors remain those of the original model; only the
autoregressive text policy was trained, without speculative decoding or MTP training.
Load a checkpoint with subfolder="iter_100" and its own tokenizer/chat template.
The subfolder preserves the original model's license, notices and README for attribution;
this root card describes the research post-training applied to those weights.