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

Assay Transfer Raw Pair V7 Intern

V7 keeps V6.5's pairing/labeling/rendering design (pipeline/v6_intern.py's single query/retrieval pairwise sigmoid target — see below) but sources its evidence from TxAgent's independently normalized, condition-key-bucketed, and endpoint-calibrated Bioavailability_Ma artifact instead of this repo's own pipeline/source_normalization/ + metrics.yaml derivation.

What changed vs V6.5.x

1. Different evidence source. Records, condition-key ("pair bucket") assignment, and endpoint-specific transfer/not-transfer thresholds all come from TxAgent's Bioavailability_Ma v5 normalized-evidence + endpoint-policy-v2 registry, not this repo's own pipeline. Benchmark-test-set exclusion is TxAgent's own Bioavailability_Ma random+scaffold splits (matched on true parent structure identity), not TDC.
2. 4 assay concepts, not 5. TxAgent's current Fg (gut-wall) normalization has zero valid records (a real upstream data-coverage gap, not a bug in this build) — V7 covers Fa, Fh, oral_bioavailability, oral_exposure only.
3. Pairs are gated on matching condition key (pair_bucket_key), not just endpoint. Two records can only be paired if they share the same source, endpoint, unit, and condition-key fields (assay system, species, dose bucket, report type — whichever apply per source). This prevents, e.g., comparing a hepatocyte-system intrinsic-clearance value against a liver-microsome value as if they were the same measurement.
4. Downsized: 1,000,000 train pairs (flat, one row per pair), 1,000 decisive comparisons each for validation/test, 10,000 each for the ranking splits (500 distinct-molecule anchors × 20 candidates).
5. Train diversity cap. No single molecule can appear as a retrieval partner more than 8 times across the whole train split (previously uncapped).
6. New is_decisive column, every split. True when the pair's distance is outside the deadband (distance <= transfer_max or distance >= not_transfer_min); False for deadband pairs. Train is 78.3% decisive / 21.7% deadband (validation/test are 100% decisive by construction). Older pipeline versions dropped deadband pairs from training entirely; V7 keeps them (for volume/diversity) but makes the distinction visible instead of silent.
7. Prompts are source-native and pipeline-agnostic. Displayed values/units come from the original reported (measurement_text, unit_text) pair, not TxAgent's cleaned (canonical_measurement, canonical_unit) — the two are not interchangeable one-for-one (canonical_unit sometimes folds scientific notation into the number), so the raw pair is used as-is rather than reconstructed. The endpoint label is the raw source endpoint_name. The transfer/not-transfer criterion is phrased as a plain chemistry statement generated from the raw threshold numbers (e.g. "within 2-fold... 5-fold or more apart") — prompts never name the upstream system or its internal policy/threshold-family taxonomy.
8. Leak fixes in the query record. Two source-text fields were found (during construction of this dataset) to sometimes restate the query's own hidden value in prose: support_text (nearly always states the row's own measured value) and, less often, comparator/extra_details (sometimes states a closely related reference value, e.g. an IV exposure used for a relative-bioavailability comparison). All three are shown only for the retrieval record (whose value is disclosed anyway) and never for the query record. A directional/orientation cue for permeability and efflux endpoints (e.g. apical-to-basolateral) is preserved as a value-free categorical label instead, since it cannot leak a magnitude.
9. metadata shape matches the SFT/ranking training contract. Every row's metadata is {"soft_targets": {"(A)": target_a, "(B)": target_b}, "groups": {"assay_concept": <top-level assay_concept>}, "target_z": target_z} — soft targets keyed by the raw completion tokens (for soft-target training) and grouping nested under groups (for grouped/ranking evaluation), not the flat target_distribution/top-level-only shape used internally by pipeline/v6_intern.py::row_for().

The eligible-records adapter and build driver are pipeline/source_normalization/starling_txagent_eligible.py and scripts/build_v7_intern_raw_pair.py.

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 — this target is scale-invariant to a common linear rescale of (comparison_value, transfer_max, not_transfer_min), so it reproduces the identical result as computing it from TxAgent's raw values.

Splits

• train — 1,000,000 flat directed pairs (one row per pair).
• validation / test — 1,000 decisive comparisons each (125 per concept × 2 labels × 4 concepts).
• validation_ranking / test_ranking — 500 distinct-molecule anchors × 20 candidates = 10,000 comparisons each.

All five splits share one schema (split-specific columns null-filled). Intern tokenization is pinned to jiosephlee/Intern-S1-mini-lm revision fcb667c380ae01f57693a45b4b5c2d331052a107.

Known limitations

• confidence and molecule_name are not included. TxAgent's normalized-evidence records carry a per-record extraction-confidence score (0.65–1.0 among valid rows, mean 0.97) and a human-readable source molecule name; neither is carried into this dataset's schema, so there is no built-in way to filter or trace rows by either.
• canonical_assay_system (part of pair_bucket_key) is fuzzy-matched for ~3.3% of valid TxAgent records (assay_system_mapping_status == "reviewed_fuzzy", TxAgent's own reviewed fuzzy-match tier, not an unreviewed guess). This is a small residual risk to the "both records in a pair share the same assay system" guarantee described above — a fuzzy match could in principle merge two genuinely different assay-system labels into one bucket. Not introduced by this build; inherited from TxAgent's upstream normalization.