datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
RAQUEL2-ICLR
RAQUEL2
Execution-grounded evaluation for machine unlearning. Each evaluation record is
a question answered by a SQL query run against two databases: one built from
the full corpus, and one with the forget-set facts removed. A record is
affected when the two databases disagree, and unaffected when they
agree — so the label is a measured property of the data, not an annotation.
Four configs, together enough to run the benchmark end to end:
Config / split
What it is
Use… See the full description on the dataset page: https://huggingface.co/datasets/Hyukkyu/RAQUEL2-ICLR.surface-audit
TruthfulQA-476 — a surface-form-cleaned binary-choice TruthfulQA
TruthfulQA-476 is the recommended drop-in replacement for the binary-choice TruthfulQA
evaluation set. It keeps 476 of the 790 original question pairs, in the original schema, chosen so
that a classifier restricted to six surface features of the answer text (negation, hedging, length,
token statistics) barely separates correct from incorrect answers (AUC 0.528, at the edge of statistical detectability), while the… See the full description on the dataset page: https://huggingface.co/datasets/iclr2027-surface-audit/surface-audit.