You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access requires manual approval by the maintainers.

Log in or Sign Up to review the conditions and access this dataset content.

Spurious-Correctness Evaluation Set v2 (BIRD dev)

1,043 execution-matched (question, gold SQL, predicted SQL) pairs from BIRD dev, adjudicated at the QUESTION level:

  • SPURIOUS (178): outputs match on the shipped database, but the prediction answers the question wrongly - a realistic database state separates it from gold.
  • NOT_SPURIOUS_EQUIVALENT (544): same answer on every realistic state.
  • NOT_SPURIOUS_DEFENSIBLE (320): a different but legitimately licensed reading of the question.

Label tiers: A-machine-certificate (execution-proven, 250), B-unanimous / B-weighted multi-judge consensus, plus Opus-consensus and Fable adjudication for hard cases. witness_pack carries concrete divergence evidence (rows that split the two queries); wrong_part marks the culprit SQL fragment. Complex fields are JSON-encoded strings.

Built for the SEMSHIELD-SQL study of spurious correctness in Text-to-SQL RL rewards.

Downloads last month
7