A BIG-bench true/false fact-checking task over Wikipedia FEVER claims, COVID-19 scientific claims, and Politifact items.
unassessed
| Category | knowledge |
|---|---|
| Subcategory | BIG-bench true/false fact-checking (FEVER, COVID-19 scientific, Politifact) |
| Page status | unknown |
| Metric | multiple_choice_grade |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 7154 |
| Dataset licence | Apache-2.0 |
| Publisher | Google (BIG-bench collaboration); HKUST authors |
fact_checker asks whether a short English claim is true or false, without retrieved evidence. Three subtasks share that format: FEVER Wikipedia claims (supports/refutes mapped to true/false; "not enough info" dropped), COVID-19 scientific myths, and Politifact political claims balanced across parties. The skill is parametric fact-checking, not evidence-based FEVER verification with Wikipedia pages. It is not [bbh](bbh.md).
Two-option multiple choice (true/false), preferred metric multiple_choice_grade. Every counted JSON item uses a dated prefix: "On <date>, the following claim was made: … Q: Was this claim true or false?" append_choices_to_input is false. Canary GUID embedded. Keywords include zero-shot, one-shot, and many-shot. The README still shows an older undated "Claim: … Is this claim true or false?" prompt that the merged JSON no longer uses.
No model card in ModelSpec reports this benchmark yet.