Sync evaluate.py with repo (reasoning-model + stratified easy/hard fixes)

Update dataset card to v3

v3: cap 1500/class, hard-weighted train split (train-only), low-info filter, 69 reviewed gold-label overrides

Update citation URL to exploitintel (username change)

Fix card: capped/balanced build (69,386 rows), accurate stats

v2: cap majority CWEs at 2500 (69,386 rows; rebalanced long tail)

Add eval script (unsloth loader for gemma4)

Add dataset card

Upload dataset

initial commit