jiosephlee/assay-transfer-record-level-v23-2-1-bbb-martins-l5-intern

BBB V23.2.1: selected continuous target policies

Balanced-sampling derivatives of the frozen V23.2 degree96 individual level datasets. L3 and L4 use threshold 1.0 SD and temperature 0.05; L5 uses 1.0 SD and 0.4. These are rank one by mean ID KNN MAE@5 across Ridge, weighted cosine and RF in the complete seed-42 molecular sweep. This selection has not been validated by an LLM comparison. The selected L3 policy worsened mean surrogate OOD MAE.

Continuous targets in both splits use sigmoid((threshold_sd - standardized_canonical_value_difference) / temperature). All target representations, completion, binary_label and decisiveness are updated. Empirical targets on binary measurement rows are unchanged.

Training pairs are regenerated around the new 1.0-SD near/far boundary, aiming for 50/50 per query where eligible partners and remaining caps permit. Near means standardized distance <= 1.0 in the same reviewed raw/log geometry used by the target. Shortages are filled from the available side. Record-role degree remains 96 and the global parent-role cap remains 576 across the original five-level schedule. L1/L2 keep their original boundary in that schedule and are not exported here.

Record train/validation membership, prompt rendering, exact assay-bucket/level intersections, UID provenance, original raw/log geometry, calibration SDs and Morgan ranking candidate pools are preserved. Binary sampling keeps its category rule; shared parent caps can affect its realized pairs. Binary probabilities remain frozen at the original V23.2 source-native empirical estimates. Calibration remains V23.2's train-plus-validation bucket calibration; no SD is refit. Gold query/reference values and labels used by KNN evaluation remain unchanged. Training ranking_balance_label reflects the new continuous sampling boundary (or the unchanged binary category match rule). Validation is Morgan-selected, not near/far balanced, and keeps its original ranking labels and candidates. There is no test split. No inclusive-bucket or degree48 rebuild is included.

calibration.json remains byte-identical to V23.2. target_policy.json records the new sigmoid; continuous-row metadata identifies the new target contract. The manifest pins the parent release, implementation, sweep and output hashes.

From the repository root, using the project's openrlhf_tfv4 Python environment:

python -m assay_transfer.record_level.v23_2_1.build

The builder archives the previous target-only release before replacing outputs. It checks sampler caps, training eligibility, bucket identity, near/far labels, and every exported target representation. The manifest includes realized balance and added/removed pair counts. Published target-only revisions remain in HF history.

Dataset: L5

Policy: {'threshold_sd': 1.0, 'temperature': 0.4, 'formula': 'sigmoid((threshold_sd - standardized_canonical_value_difference) / temperature)', 'applies_to': 'continuous rows in train and validation_ranking', 'selection': 'rank 1 by equal-model mean ID KNN MAE@5, complete seed42 sweep', 'binary_targets': 'unchanged', 'geometry_and_calibration': 'unchanged', 'sampling_policy': {'binary_probabilities': 'frozen V23.2 source-native empirical targets', 'continuous_boundary_sd': 1.0, 'global_parent_role_cap': 576, 'global_schedule': 'all five frozen V23.2 levels; L1/L2 retain 0.4 boundary', 'near_rule': 'standardized_logit_distance <= 1.0', 'record_role_degree_cap': 96, 'shortage_policy': 'fill from available side', 'target_near_fraction_per_query': 0.5, 'validation_selection': 'unchanged Morgan ranking pools'}}

Rows: {'train': {'rows': 9727, 'continuous_rows': 4320, 'changed_targets': 4320, 'changed_completions': 1102}, 'validation_ranking': {'rows': 270, 'continuous_rows': 210, 'changed_targets': 210, 'changed_completions': 48}}