split_protocol.md

Split protocol

Train, validation, and test partitions use 60%, 10%, and 30% of unique paragraph groups, respectively. The deposited assignment CSVs are the authoritative allocation and remove any dependence on library-specific random shuffling.

All rows sharing a Paragraph_ID stay together. Since every Tuple_ID belongs to a single paragraph, this also keeps every P1–P5 tuple family together and prevents tuple-family leakage. Reddit extraction additionally groups by Document_ID where that identifier is available. src/validation/validate_splits.py checks these invariants across all released partitions.

src/preprocessing/materialize_probe_splits.py rebuilds probe partitions directly from a canonical probe table and the deposited assignments. Example:

python src/preprocessing/materialize_probe_splits.py \
  --probes data/probes/pubmed_probes.csv \
  --assignments data/splits/probes/pubmed/split_assignments.csv \
  --output-dir /tmp/pubmed_splits

Synthetic assignments are in data/splits/probes/synthetic/exp3_2_probe_split_assignments.csv; PubMed and Reddit use split_assignments.csv in their domain directories.