Fifth revision: the blind human read

Third and fourth revisions: held-out evaluation and measure 3b

Correct all figures against trainer_state.json; restore 364,831 denominator

Restore YAML metadata and markdown; correct step figures to trainer-recorded values; add evaluation status

Correct model card: research-checkpoint status, remove unsupported metrics

Upload README.md with huggingface_hub

Upload folder using huggingface_hub

initial commit