Update to 1-epoch SFT model (label fix + response masking): better generalization

learnllm-100m-instruct: SFT model (response-masked attention + label fix)

initial commit