Okapi multilingual TruthfulQA

lm-eval group of GPT-translated TruthfulQA val items in 31 languages, scored as MC1 accuracy and MC2 probability mass.

Also known as: truthfulqa_multilingual, m_truthfulqa, alexandrainst/m_truthfulqa

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.
Categorysafety
Subcategorymachine-translated multiple-choice truthfulness
Page statusactive
MetricMC1 accuracy; MC2 normalised true-answer probability (acc)
Directionhigher_is_better
Dataset size23586
Dataset licenceCC-BY-NC-4.0
PublisherUniversity of Oregon NLP (translations); Alexandra Institute (Hub dump); EleutherAI (lm-eval group)

What it measures

okapi_truthfulqa_multilingual is EleutherAI's group over machine-translated [TruthfulQA](truthfulqa.md). The model is asked a question written to bait a popular misconception and must prefer the true continuation over false ones. Hub dump alexandrainst/m_truthfulqa was translated with GPT-3.5-turbo by University of Oregon and first lived in nlp-uoregon/mlmm-evaluation. The Okapi paper's own eval suite is ARC, HellaSwag, and MMLU; this TruthfulQA translation uses the same pipeline but is not one of those three paper tables. English text only in the six-shot prefix; questions are in the target language.

Task format

Multiple choice, two YAML variants per language: truthfulqa_{lang}_mc1 and truthfulqa_{lang}_mc2. README group/tag truthfulqa_multilingual. MC1 target index is 0 (first listed choice). MC2 uses labelled true/false options and sums normalised probability mass on true labels. Validation split val. process_docs prepends a fixed English 6-QA prompt before the translated question.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub