RESTORE v2's actual training build (overwritten in error 2026-07-15) + rewrite the card to document its defects. The corrected build ships as v4.

rebuild 2026-07-15: fix vocab-tail truncation (caps were prefixes, not samples) + add grammatical single_poison [card]

rebuild 2026-07-15: fix vocab-tail truncation (caps were prefixes, not samples) + add grammatical single_poison [n50]

rebuild 2026-07-15: fix vocab-tail truncation (caps were prefixes, not samples) + add grammatical single_poison [n100]

n50/eval_unseen.jsonl: v2 (=v11) coverage build

n50/eval_seen.jsonl: v2 (=v11) coverage build

n50/eval_natural.jsonl: v2 (=v11) coverage build

n50/dev.jsonl: v2 (=v11) coverage build

n50/train.jsonl: v2 (=v11) coverage build

n100/eval_unseen.jsonl: v2 (=v11) coverage build

n100/eval_seen.jsonl: v2 (=v11) coverage build

n100/eval_natural.jsonl: v2 (=v11) coverage build

n100/dev.jsonl: v2 (=v11) coverage build

n100/train.jsonl: v2 (=v11) coverage build

Upload TRIGGERS.md with huggingface_hub

Upload triggers.json with huggingface_hub

v2 dataset card

initial commit