jiosephlee/assay-transfer-raw-pair-v7.1-intern

Assay Transfer Raw Pair V7.1 Intern

V7.1 is a thin, additive update to V7: same pairing/labeling/rendering design (pipeline/v6_intern.py), same TxAgent-sourced Bioavailability_Ma evidence, but built after TxAgent rebuilt its Fg (gut-wall availability) normalization, and after several diversity/leakage fixes to the train-generation logic itself. Nothing about V7's four original concepts' records changed between the two eligible-records snapshots — this is purely additive plus a train-generation rework.

What changed vs V7

1. 5th assay concept: Fg. TxAgent's Fg normalization previously produced zero usable records (V7 covered Fa, Fh, oral_bioavailability, oral_exposure only). The rebuilt snapshot yields 1,056 eligible Fg records (462 distinct molecules, gut-wall efflux/secretory-transport assays — Caco-2/MDCK bidirectional permeability, P-gp/BCRP substrate classification). Fg is train-only: it never appears in validation, test, or either ranking split.
2. Validation/test are frozen to V7's exact held-out molecules, not re-derived from the new (larger, Fg-inclusive) weight distribution. scripts/build_v7_1_intern_raw_pair.py's --split-mode reuse (the default used for this build) recomputes V7's original 1,250+1,250 validation/test molecules deterministically from its untouched eligible-records artifact, so re-deriving a fresh held-out set here can't silently drift the eval population just because Fg added new record weight to the pool. (--split-mode recompute is also available, for deriving a fresh held-out set from this dataset's own distribution instead.)
3. No cross-split molecule leakage, including for Fg. A new module, pipeline/split_leakage.py, checks any "force this concept into that split" override (here: routing Fg into train) against the base partition and drops exactly the rows that would leak a molecule into two splits, rather than allowing it. 108 of the 1,056 Fg records belong to molecules already held out in V7's validation/test and are dropped outright (not moved, not usable anywhere); the remaining 948 Fg records — including 148 on molecules that don't exist in V7 at all — contribute to train.
4. Train anchor-frequency skew fixed. V7 had no cap on how often a record could be selected as the query side of a training pair (only the retrieval side was capped at 8) — a heavily-studied molecule with thousands of raw records could dominate query selection by tens of thousands of pairs. TRAIN_QUERY_DEGREE_CAP now bounds this too. Both this and the pre-existing retrieval cap are 6 (each record: at most 6 as query + 6 as retrieval = 12 total appearances), keyed by raw record id, not canonical SMILES (a molecule with many raw records is not deduplicated across them).
5. Pair-bucket diversity. Query selection now round-robins two levels deep — concept, then pair_bucket_key within that concept, then record within that bucket — instead of a flat per-concept record list, so a mega-bucket with thousands of records no longer gets proportionally thousands of times more query-attempt volume than a small one.
6. Small buckets can now contribute training pairs. V7 discarded any attempt that didn't produce exactly 4 chosen retrievals, which structurally excluded every pair_bucket_key with fewer than 5 total records (over half of all pairable buckets). Groups can now be 1-4 members wide in the default flat build (--listnet mode is unaffected and keeps the strict requirement, since ListNet's ranking loss assumes a fixed width).
7. Train volume is a soft target, not a guarantee. Because of the frequency caps above, generation now naturally exhausts (every record in every reachable bucket capped or structurally out of partners) before necessarily reaching the requested pair count. This build requested 1,250,000 flat train pairs and produced 952,346 — build() does not error on falling short, only on producing zero usable pairs.
8. New templates/assay_transfer_v7_intern/Fg.jinja, matching the existing per-concept template style (see Fa.jinja/Fh.jinja) and using only the context fields TxAgent's Fg schema actually populates (measured process, direction, assay system, intestinal site, transporter/enzyme, substrate status, qualifying conditions, extra details, support text).

Everything else — the sigmoid target, pair_bucket_key condition-key pairing gate, is_decisive column, source-native prompt rendering, query-record leak fixes, and metadata shape — is unchanged from V7 (see its README for details).

The eligible-records adapter and build driver are pipeline/source_normalization/starling_txagent_eligible.py and scripts/build_v7_1_intern_raw_pair.py (a thin wrapper importing V7's build script directly, not a copy).

Target

d = abs(comparison_value_retrieval - comparison_value_query)
z = log(9) * (transfer_max + not_transfer_min - 2*d) / (not_transfer_min - transfer_max)
q_A = eps + (1 - 2*eps) * sigmoid(z / T)      # eps = 0.1, T = 1 (label smoothing)
q_B = 1 - q_A

comparison_value/transfer_max/not_transfer_min are all in TxAgent's own normalized (0–100) distance space rather than raw physical units.

Splits

• train — 952,346 flat directed pairs (one row per pair; requested 1,250,000, see point 7 above). All 5 concepts, including Fg.
• validation / test — 1,000 decisive comparisons each (125 per concept × 2 labels × 4 concepts — Fg excluded).
• validation_ranking / test_ranking — 500 distinct-molecule anchors × 20 candidates = 10,000 comparisons each (Fg excluded).

All five splits share one schema (split-specific columns null-filled). Zero molecule overlap between train and validation/test, verified directly against the built parquet files. Intern tokenization is pinned to jiosephlee/Intern-S1-mini-lm revision fcb667c380ae01f57693a45b4b5c2d331052a107.

Known limitations

Inherited from V7 (see its README): confidence/molecule_name not included; canonical_assay_system is fuzzy-matched for ~3.3% of valid TxAgent records.

Specific to V7.1: under --split-mode recompute, a molecule whose only eligible records are Fg can be selected into the fresh validation/test held-out set purely on Fg record weight, then lose all its rows once Fg is excluded from eval — so that mode's held-out molecule count can land slightly under the 1,250/1,250 target (observed ~1,238/1,241). This build used --split-mode reuse (the default), which does not have this issue.