| # Dataset Protocol | |
| This project locks data protocol before model training. | |
| ## ViFactCheck | |
| Main protocol uses `Statement` plus retrieved chunks from `Context`. The gold `Evidence` field is diagnostic only: upper-bound, recall/rationale checks, and audit support. It must not be used as main verifier input. | |
| ## AVeriTeC | |
| The local official `test.json` is hidden-style and does not contain labels or QA evidence. It is prediction-only unless official evaluation labels are available. Local evaluation uses official dev as `local_test`; official train is deduplicated against official dev by normalized claim hash, then group-split into `train_inner` and `dev_inner`. | |
| ## HealthVer | |
| Main unit is pair-level `claim + evidence -> label`, matching the dataset structure. Claim-level grouped files are produced for robustness and aggregation analysis only. | |