Add quality-filtered validation and test splits to filtered-full
Adds quality-filtered validation and test splits to the existing filtered-full configuration.
| Split | Original | Keep | Uncertain, retained | Rejected | Published |
|---|---|---|---|---|---|
| validation | 15,000 | 13,640 | 1,091 | 269 | 14,731 |
| test | 15,000 | 13,561 | 1,184 | 255 | 14,745 |
Both derive from the same frozen full revision as filtered-full/train: d2ab1d12be4463fb7ac1f877ed2193b3bb59642a. GLM-5.3-Flash reviews all 30,000 rows using the existing conservative two-reviewer and retention-guard policy. A follow-up final challenge audit checks all proposed exclusions and retains defensible targets; its protocol and one conservative spot-check retention are documented separately. Uncertain cases are retained. No downstream decider/JEV predictions are used. All primary reviews completed successfully.
The held-out selection is quality-only: it does not apply the training-only source, commercial-license, or length exclusions. Existing license metadata and all original row values/soft targets are preserved. Original default/full data and filtered-full/train remain unchanged. Subsequent source additions are outside this historical snapshot.
Includes a per-row decision ledger, machine-readable counts/provenance, and methods documenting selection effects and limitations. Validation checks frozen inputs, every exported row after Parquet round-trip, split metadata, and local dataset loading. Uploaded hashes and loading from the PR revision are verified after creation.