jiosephlee's picture
Publish original Oral V25 mixed L2 L3 L4 L6, excluding L5
0c193e6 verified
|
Raw History Blame Contribute Delete
2.15 kB
metadata
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

Oral bioavailability V25: mixed without L5

Independent L2-L6 releases derived from the frozen consolidated oral V24.1 splits. Record degree is 24 for L5 and six for L2/L3/L4/L6, separately for query and reference roles. Parent molecule budgets remain 4096 per role, level, and binary/nonbinary family. Minority-first sampling, compatible-endpoint repair, and final binary 75/25 trimming are retained. Low degree and binary trimming can leave records unused; see record_coverage.json.

The entire official validation and test connected scaffold universes are excluded from training across all levels, including records not selected as evaluation queries. Evaluation membership and reference ordering match oral V24.1: 500 total queries per level per split, Morgan top-50 ID references, and complete OOD pools.

Conflict exclusions, support thresholds, reviewed identity/log10 geometry, and the ratio gate <=0.5 (undefined rejected) are preserved. Test measurements only participate in the inherited eligibility gate, not training target fitting. Numeric targets retain center 0.4 SD and temperature 0.1. Binary soft targets are refitted globally within each level from the final numeric training pairs, and all training, validation, and test rows are rendered using those updated targets. Internal v24.2 sampler/schema names identify reused algorithms, not the release.

Mixed L2/L3/L4/L6, excluding L5

This release concatenates original V25 pairs within each split in L2/L3/L4/L6 order. No cross-level pairs, resampling, target refitting, or prompt edits are performed. All fields and per-level degree-six caps are preserved exactly. Each evaluation split contains 2,000 queries (500 per included level). V25.1 cleanup and canonical-display changes are NOT included: original V25 prompts can lack explicit units. The identical source calibration file is retained, including unused entries for excluded levels.