metr-evals/

malt-transcripts-public

MALT-public is our collection of agent transcripts. Our public variant only includes data on non-internal tasks, which includes 30 task families and 169 tasks, across ~19 different models (some might be different releases of the same model, from different providers, or internal naming changes). Here's a summary table:

has_chain_of_thought labels model manually_reviewed run_source count

False bypass_constraints… See the full description on the dataset page: https://huggingface.co/datasets/metr-evals/malt-transcripts-public.