laya-finance-matched-sft-continued-200k

Paper: S²-MAS. The fixed-budget adaptation results and matched downstream methods define the comparison reported here.

Project background

S²-MAS is a society-scale multi-agent simulation framework for studying how household decisions affect consumption, cash reserves and payments. The paper connects a constructed resource of 500 million household-persona records, learned decision policies, reusable financial scenarios and a state-transition engine. Its purchase-funding application evaluates 12 million distinct synthetic household agents under paired conditions. Full-data Fin-Laya supplies the policy for that application; this repository supplies a fixed-budget adaptation checkpoint for the common decision-model comparison. Decision learning uses real-survey data. The GENESE/MAS engine implementation remains private.

The benchmark groups twelve systems into six decision-model routes: encoder option scoring; language models with learned decision heads; next-token decision SFT; dynamic-schema classification; multimodal encoder decisions; and provider decision interfaces. All receive text inputs in the common benchmark. The 17 model-repository entries comprise 10 fixed-budget trained variants—five native Laya and five Qwen variants—plus six checkpoint/runtime entries and one compatibility alias.

This model's role

This checkpoint is the cross-entropy continuation control within the encoder option-scoring route. It starts from matched SFT200 and rereads the same 200k examples, sharing the parent and added full-pass budget with RLCD200. Development selects the checkpoint after 50,688 additional presentations, giving 250,688 selected cumulative exposures.

Training design and checkpoint

Method: Cross-entropy continuation from matched SFT200.
Distinct downstream training rows: 200,000 from the fixed decision-QA training partition.
Initialization: ya9919/laya-finance-matched-sft-200k.
This stage terminal row exposures: 200,000; selected checkpoint stage exposures: 50,688.
Selected checkpoint cumulative row exposures: 250,688; selected cumulative distinct rows: 200,000.
A fixed 20,000-example development set selects this checkpoint.

Performance and matched comparisons

Language: English. Inference precision: Native PyTorch BF16.

The common held-out test contains 240,458 QA records, 119 question versions and 8,275 respondent/household groups. The checkpoint is selected using development data before this test.

Checkpoint / reference	Accuracy (%)	Question-macro F1 (%)
This checkpoint	73.8986	36.1732
SFT200 parent	73.8753	35.9949
Original native Laya, matched BF16 reference	60.1972	29.5995

Accuracy weights every test record equally; invalid decisions count as incorrect. Question-macro F1 averages over every defined candidate class within each question version and then weights the 119 versions equally.

Against its SFT200 parent, the paper reports +0.0233 accuracy points (95% CI −0.0288–0.0731) and +0.1783 question-macro F1 points (0.0762–0.2770). The CE continuation has 0.1229 F1 points more than RLCD200 under the equal added-pass budget (95% CI 0.0330–0.2073).

Package and inference

Complete fine-tuned encoder and original native decision head.

This stage continues the matched SFT checkpoint for one pass over the same training records.

Reported common-test inference: native BF16.
Underlying encoder: answerdotai/ModernBERT-large; the original two-layer Laya decision head is retained.
Inference scores the full state, question and complete candidate menu.

Use the native Laya inference interface with the supplied encoder configuration, tokenizer and decision head.

License and access

The project model contribution is offered under the S²-MAS Controlled Model Research License v1.2: noncommercial scientific use, no redistribution, and no unauthorized reverse engineering, extraction or distillation. Training and evaluation records are not distributed. Source-data access requires the provider's authorization. Permission requests: yli9919@hku.hk.

Project repository: s2-mas.