Remove obsolete WikiText-only reconstruction and unused reference-dev files. Put fast_validation inside stage-2, rename English reconstruction to stage-1, and retain all selected data bytes.
Keep combined data and remove redundant source configs
Remove 12 per-dataset configs and their duplicate partitions. The default train/validation/test data continues to identify sources in the dataset field.
Update dataset files and use standard dataset names
Overwrite original train/validation/test/reference/fast files with the verified 12-dataset data. Remove MCTest, replace ad hoc config names, and remove the superseded separate upload. Reconstruction pretraining data preserved.
21 days ago
Add verified situational-v1 dataset and reproducible splits
Frozen training and reporting data used by the current LatComm runs; original JSONL bytes preserved in gzip, plus lossless Parquet views, source metadata, split membership and checksums.
