This is the fully merged checkpoint obtained by AsFT fine-tuning wvnvwn/llama2-7b-chat-lr5e-5-ssft-cb on the MedQA training split.

Training used 10,176 retained examples for 3 epochs with learning rate 3e-4, physical and effective batch size 16, cosine scheduling with warmup ratio 0.1, BF16, maximum sequence length 1,024, and seed 42. Two examples with no response tokens after preprocessing were excluded. The LoRA configuration used rank 16, alpha 32, dropout 0.05, and target modules q_proj, k_proj, v_proj, up_proj, and down_proj. AsFT used 160 alignment directions and lambda_reg=1.0.

The Llama-2 chat template was applied and prompt tokens were masked from the training loss. The saved weights have already been merged into the starting model; no separate adapter is required for inference.