---
configs:
- config_name: default
  data_files:
  - split: train
    path: train/data.parquet
  - split: validation_ranking
    path: validation_ranking/data.parquet
  - split: test_ranking
    path: test_ranking/data.parquet
---
# BBB all-record assay transfer V21

V21 freezes V20 records, splits, source-native prompts, target geometry, and 20-parent
Morgan ranking panels. Training buckets are ranked by train-record count and assigned
independent per-record query/retrieval degree caps of 96, 48, 24, 12, or 6 by size
quintile. Direct-BBB records always use cap 6, and deterministic bucket-balanced
downsampling limits direct pairs to 50% of training rows. Parent roles remain capped at 576.

- train rows: 480,974
- validation ranking rows: 2,800
- test ranking rows: 5,060
- achieved direct-BBB train share: 0.500000
- continuous target: unchanged V20 train-SD sigmoid centered at 0.4 SD
- prompts and evidence: original source values and units
- log10 geometry: machine-readable metric geometry only, never prompt text