Add filtered-full train configuration (1,305,478 retained decisions)
Adds a filtered-full configuration with a single train split containing 1,305,478 retained decisions from full/train at revision d2ab1d12be4463fb7ac1f877ed2193b3bb59642a.
The selection combines existing license-metadata, benchmark-overlap, structural, and token-budget filters with conservative GLM-5.3-Flash review through llm-api. Of 2,500,000 source rows, 1,328,623 were eligible for review; 23,138 were rejected and seven unresolved API-error rows were excluded. The retained set includes 97,030 uncertain cases. This is a filtered historical snapshot and does not include subsequent repository additions.
Original row values, option order, decision kinds, soft targets, IDs, and license metadata are preserved in the 13-column schema of that source revision. The newer example_id column is not synthesized. Existing default/full files and validation/test splits are unchanged.
The added dataset documentation explains the filtering policy, review reuse, conservative retention overrides, limitations, and partially retained groups. A machine-readable report records counts and checksums.
Validation: source and selection fingerprints checked; all retained IDs and contents matched; every exported Parquet value read back and verified; dataset-card metadata validated; local Hugging Face Datasets streaming load checked. The PR branch is also checked for uploaded file hashes and a working load_dataset(..., "filtered-full", split="train", revision="refs/pr/N") load after creation.