README.md

---
configs:
- config_name: default
  data_files:
  - split: train
    path: train/data.parquet
  - split: validation_ranking
    path: validation_ranking/data.parquet
---
# BBB assay transfer V23.2: UID level datasets

Five separate, noncumulative datasets use BBB evidence library V10 and the pinned
upstream acquisition UID mapping. L1 = direct brain exposure; L2 = central functional
access proxy; L3 = passive permeability; L4 = efflux transport; L5 = influx transport.
Both the query and reference belong to the dataset's level. Source identity is
preserved independently. The mapping is joined on `source_row_uid`, never on
SMILES, source IDs, or upstream canonical IDs. Missing UIDs are excluded with an
identity audit; they are not assigned an inferred level.

V23.2 rebuilds from V10. It is not a filter of the published V23/V9 pairs.
Each exact V10 assay bucket is intersected with level before calibration, split
support gates, pair sampling, and ranking selection. `upstream_pair_bucket_key`
preserves the original bucket; `pair_bucket_key` appends the level to its JSON key.
Both acquisition UIDs and immutable canonical record IDs are retained in release rows.

V23's relaxed initial calibration gate (10 records, no initial molecule minimum)
is reconstructed using the V10 calibration implementation. Subsequent training
requires 12 training records and 6 distinct training parents per level-bucket.
Whole level-buckets with 10–11 records and at least 6 parents form OOD validation.
ID holds out the gold-validation scaffold components across all levels together.
OOD is bucket OOD; parent overlap with other buckets is possible.

Continuous targets retain V23's reviewed geometry, with recalculated sample SD
(ddof=1) on train plus validation within the level-bucket, or the whole OOD bucket.
Original assay keys select reviewed log transforms. Unreviewed tail alerts and
invalid transformations are excluded and audited. Binary calibration requires
both categories with at least three records each. Empirical binary targets retain
V23's source-native estimator, fitted only on retained training pairs.

The nonmixed build gives every level-bucket a record-role degree cap of 96,
including L1. It exports five separate L1-L5 datasets. The mixed build pools
all levels in one dataset and uses degree 48 for buckets with at least the
configured number of training records, and degree 96 for smaller buckets.
By default the cutoff is the smallest integer above the geometric mean of
training level-bucket sizes: 33 records for this pinned split (355 of 944
training buckets). Override with `--big-bucket-min-records` when building pairs.
Neither variant applies an L1-specific cap or share ceiling. Both retain the
global parent-role cap of 576 across levels. Degree is a cap, not a guarantee:
limited eligible partners and parent caps can reduce the realized degree.
The two variants are sampled separately and have separately fitted empirical
binary targets. Mixing datasets does not permit cross-level pairs.

Validation bucket quotas are computed across levels separately for each regime
and measurement family, using V23's second-smallest positive count rule. Up to
three heldout queries are selected per bucket. ID uses 19–20 nearest training
records; OOD uses the other 9–10 records. Ranking groups are complete and stay
within one level. Sparse level panels retain their actual support. There is no
test split. `assay_concept` and the metadata level group identify the level;
`source_id` continues to identify scientific source provenance.

Prompt rendering and masking reuse V23. Level and UID fields are provenance,
not extra prompt evidence. Each dataset carries the complete calibration artifact
because rows refer to its SHA-256. Input hashes are pinned in `inputs.json`.

Use the project `openrlhf_tfv4` Python environment:

```bash
python -m assay_transfer.record_level.v23_2.build all --variant all
python -m assay_transfer.record_level.v23_2.build verify
```

Outputs: `assay_transfer/record_level/artifacts/v23_2/degree96/hf/bbb_martins/by_level/L1`
through `L5`, plus `degree96/hf/bbb_martins/mixed`. The verified V10 splits and UID audit are shared with the original build.
`all` rebuilds pairs and exports; `splits` is a separate explicit reconstruction.
Use `--variant nonmixed` or `--variant mixed` for one variant. Existing outputs
require explicit `--overwrite`. The original lower-degree artifacts are retained.

Dataset: L5

Rows: {'train': 9723, 'validation_ranking': 270}