Embedded Files
Hidden-RAD2 uses separate evaluation frameworks for its two tasks.
Task 1 is evaluated as structured prediction over the A1–A5 answers. Task 2 is evaluated as a combination of validity classification, error-span detection, hallucination-type classification, correction quality, and confidence calibration.
For details, see:
The official normalization rules, output format, evaluation scripts, and tie-breaking procedure will be released before the formal run.
Page updated
Google Sites
Report abuse