Datasetpublic

danielfein/raid-neologism-table-splits

RAID neologism table splits Source-disjoint RAID train-derived paired splits for the two-token AI detector experiments. Source dataset: liamdugan/raid, config raid, split train. Seed: 20260501. Base source partition: 10000 train source_ids, 3000 test source_ids, overlap 0. Row format: one human text and one same-source_id AI text per row. Protocols: standard_train, standard_test: model, attack, decoding, repetition penalty, and domain sampled randomly. model_<model>_train… See the full description on the dataset page: https://huggingface.co/datasets/danielfein/raid-neologism-table-splits.

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes69downloads

Dataset Card

RAID neologism table splits

Source-disjoint RAID train-derived paired splits for the two-token AI detector experiments.

—Source dataset: liamdugan/raid, config raid, split train.
—Seed: 20260501.
—Base source partition: 10000 train source_ids, 3000 test source_ids, overlap 0.
—Row format: one human text and one same-source_id AI text per row.

Protocols:

—standard_train, standard_test: model, attack, decoding, repetition penalty, and domain sampled randomly.
—model_<model>_train, model_<model>_test: everything random except generator; train excludes the held-out model and test uses only that model.
—attack_train: clean/standard training with attack=none.
—attack_<attack>_test: one test split per attack; everything other than attack is sampled randomly.

See metadata.json in the repository for counts and exact split semantics.