This is a fully merged WSR-LoRA checkpoint trained on the MedQA training split from wvnvwn/llama2-7b-chat-lr5e-5-ssft-cb.

Training used learning rate 2e-4, 3 epochs, physical/effective batch size 16, cosine scheduling with warmup ratio 0.1, BF16, maximum sequence length 1,024, and seed 42. WSR-LoRA used rank 16, alpha 16, dropout 0, and target modules q_proj, k_proj, v_proj, up_proj, and down_proj. Rotation reused a 160-layer safety SVD basis constructed from all 4,994 Circuit Breaker examples. Factor importance used 512 safety examples, with the top 10% of entries frozen in both A and B factors.

The weights are already merged into the starting model. See wsrlora_run_config.json for the full run provenance.