File size: 1,685 Bytes
69be449 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 | ---
configs:
- config_name: default
data_files:
- split: train
path: train/data.parquet
- split: validation_ranking
path: validation_ranking/data.parquet
---
# BBB V23.2.3: temperature0.5 with empirical binary targets
L3–L5 use threshold1.0 SD and temperature0.5. Continuous probability is
`sigmoid((1 - standardized_canonical_value_difference) / 0.5)`.
Binary match probability is the mean continuous near-pair probability; mismatch
probability is the mean continuous far-pair probability, separately by source.
Near means distance<=1 SD. The fit pools realized continuous TRAIN pairs across
the balanced L3/L4/L5 releases, all under the same1.0/0.5 policy. No validation
rows enter this fit. This replaces the previous frozen all-five-level0.4/0.1
binary estimates. Counts and fitting-input hashes are in calibration.json.
Both continuous and binary soft targets are updated in training and validation.
Pair identities/order, degree96 balanced sampling, global parent caps, prompts,
record splits, UID/level buckets, reviewed raw/log geometry and SDs, hard labels,
gold query/reference values and ranking candidate pools remain unchanged.
Calibration.json changes only its empirical_binary_targets table; all rows
reference its new hash. Validation soft-target losses therefore change, while
the fixed gold values for KNN MAE and categorical F1 remain unchanged.
This is a requested ablation, not a sweep winner. Previous versions remain intact.
Reproduce with `python -m assay_transfer.record_level.v23_2_3.build` from the
repository root using openrlhf_tfv4. Full-row assertions verify each output.
Dataset: L5
Rows: {'train': 9727, 'validation_ranking': 270}
|