ya9919/fin-laya-decision-qa

Financial decision-QA supervision

The study uses 4,172,421 underlying financial-survey observations, represented in bilingual views. Per language, the grouped partitions contain 2,925,263 training, 412,873 development, 415,688 calibration and 418,597 test examples. The two language views represent the same underlying observations.

A common model comparison uses 240,458 records, 119 question versions and 8,275 respondent groups. Accuracy counts invalid answers as incorrect. Question-macro F1 averages across the complete candidate classes within each question version and then gives each question version equal weight. Each run uses its development-selected checkpoint; the ten formal trained versions are separate from historical checkpoint/runtime entries and the compatibility alias.

The observations include recorded responses and coded financial states. Cleaning, reviewed translation and target-field exclusion produce learning examples while preserving question and candidate semantics. These are survey-derived observations, held under restricted access with applicable source terms and attribution obligations.

Current public resources

The v36 manuscript and public research materials describe S²-MAS, its 500M household persona resource, 12M-household T3 execution and separate matched monthly shock extension. Current public synthetic materials are distinct from this survey-derived training corpus:

• Independent 2M household companion, including 32,000 unlabeled task inputs from 2,000 households × 8 tasks × 2 conditions.
• 320 independently synthetic financial examples from 20 households × 8 tasks × 2 conditions.
• Wikidata context, with component-specific CC0 terms.

S²-MAS Models · S²-MAS Datasets. Research collaboration: yli9919@hku.hk.