Real or Fake Text (RoFT, BIG-bench)

BIG-bench RoFT task: a 10-sentence English passage starts human-written; pick the first machine-generated sentence.

Also known as: RoFT, Real or Fake Text

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorygeneration
SubcategoryBIG-bench multiple-choice: find the sentence where a passage switches from human text to generated text
Page statusunknown
Metricmultiple_choice_grade
Directionhigher_is_better
Dataset size15088
Dataset licenceApache-2.0
PublisherUniversity of Pennsylvania (via the BIG-bench collaboration)

What it measures

real_or_fake_text (RoFT) shows ten English sentences. The first is always human. At a random later sentence the rest become model-generated, or, in the easy subtask, sentences copied from another human article in the same genre. The model must name the first machine-generated sentence, without being told which generator was used. This is boundary detection in a hybrid passage, not a binary real/fake label. Subtasks: easy, gpt2, gpt2_xl. Not in [bbh](bbh.md).

Task format

JSON multiple choice over the passage's own sentences. Preferred metric multiple_choice_grade. Task prefix matches the human RoFT instructions. Dummy-model header: 15,088 multiple-choice queries. Parent task.json is a stub; examples live in easy/, gpt2/, and gpt2_xl/. Canary GUID embedded.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub