Render bare train queries in the benchmark's test shape; columns id, query, response, scoring_context first

Give train queries their benchmark's reply-format block

Point the card at the AutoDataSci collection

Add ground_truth, response and scoring_context beside every existing column; add the all config

Add an all config that concatenates the four tasks

Say that the fs rule errs toward removing more

Correct the duplicate-row count and what source_file is for

Add upstream file provenance; harden the equivalence key

Record the duplicate rows upstream leaves in the fs training split

Say where each task keeps its answer

Upload folder using huggingface_hub

initial commit