Full accuracy+F1 metrics: laya v2 zero-shot, laya v2 fine-tuned, setfit (95 test rows)

Pass HF_TOKEN to snapshot_download calls (authenticated resolve path)

Full-metric eval script for v2: laya v2 zero-shot, laya v2 fine-tuned, setfit on the 95-row split

Add evaluation metrics

Laya v2 (climate-context instructions) fine-tuned on GIZ vulnerability data: macro-F1 0.704 (zero-shot 0.163)

Fix UnboundLocalError: literal {label} inside write_readme f-string; reword v2-change sentence

v2: label-aware climate-context instructions; otherwise identical RLCD recipe to v1

initial commit