Rebuild swap_complexes: drop identical pairs, identity floor 0.7, rebalance val/test to train difficulty (gentle cap, no chain leakage)

Rebuild swap_complexes: drop identical pairs, identity floor 0.7, rebalance val/test to train difficulty (gentle cap, no chain leakage)

v3.7: unify preference_pairs (combine within+cross); leak-free chain-level verified

v3.6: rescue all 72,510 swap_complexes rows via cluster co-occurrence components; exact 80/10/10 split, leak-free

v3.5: combine within+cross into unified swap_complexes & swap_relationships; reseed splits (cluster 75/12.5/12.5, seed 0089) for ~80/10/10 row split

v3.5: combine within+cross into unified swap_complexes & swap_relationships; reseed splits (cluster 75/12.5/12.5, seed 0089) for ~80/10/10 row split

v3.4: drop same-species rows and TimeTree no-data pairs; refresh splits

Make train/val/test proper HF splits: per-split parquets + YAML data_files

Make train/val/test proper HF splits: per-split parquets + YAML data_files

add train/val/test split column to L3 + L4 (mmseqs2 0.3 PID cluster, 80/10/10)

add train/val/test split column to L3 + L4 (mmseqs2 0.3 PID cluster, 80/10/10)

add train/val/test split column to L3 + L4 (mmseqs2 0.3 PID cluster, 80/10/10)

filter L2/L3/L4 to seq_identity_A>=0.8 AND seq_identity_B>=0.8

v3 final: 5-layer (natives / L2 swap_relationships / L3 swap_complexes wjiaqi/evo-style / L4 preference_pairs). 3 identity metrics, species+gene labels, within-source + cross-source variants.

v3 L2 + L3: swap_relationships (79,241) + preference_pairs (316,964) with seq_identity_A/B + TimeTree divergence_mya labels

natives-filtered: drop 124 conglab rows with unmappable NCBI genus (1 Aeorestes + 123 na:na:na lineages); 50,481 -> 50,357

v3 — two L1 configs: natives (85,554, no species filter) + natives-filtered (50,481, Opisthokonta only)

v3 — L1 natives only (ortholog-filtered, Opisthokonta); drop v2 L2/L3 (broken same-organism random swaps); L2/L3 to follow when STRING Phase 1 finishes

Restructure as 3-layer dataset: natives + swap_relationships + preference_pairs (afcdb-final + conglab)

Add combined parquet (afcdb 142,598 + conglab 63,612 = 206,210 rows)