qwen3.5-9b-finance-lora-head-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 follows the language-model plus learned decision-head route. It applies the development-selected LoRA method to the independently initialized, official post-trained Qwen3.5-9B backbone. The 200k one-pass protocol permits a direct comparison with the adapted 4B model; the upstream training histories also differ between those initializations.

Training design and checkpoint

Method: lora.
Distinct downstream training rows: 200,000 from the fixed decision-QA training partition.
Initialization: Qwen/Qwen3.5-9B.
This stage terminal row exposures: 200,000; selected checkpoint stage exposures: 200,000.
Selected checkpoint cumulative row exposures: 200,000; selected cumulative distinct rows: 200,000.
A fixed 20,000-example development set selects this checkpoint. The complete development partition (412,873 examples) selected LoRA on Qwen4B; the method was fixed before the independently initialized 9B run. The final test did not select budgets or methods.

Performance and matched comparisons

Language: English. Inference precision: BF16 Qwen backbone with FP32 candidate scoring.

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 (%)	Invalid decisions
This checkpoint	74.0919	36.6622	0
Qwen4B-Base LoRA + head, same 200k protocol	74.1435	36.7343	0

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.

The paper reports 9B minus 4B = −0.0516 accuracy points (95% CI −0.0981–−0.0071) and −0.0721 question-macro F1 points (−0.1725–0.0404). This comparison covers the independently initialized backbones under the matched downstream protocol.

Package and inference

LoRA adapter plus a shared scalar decision head; obtain the official backbone separately.

Initialized from the official post-trained 9B release. The 4B/9B comparison varies both model scale and upstream training history.

Model files: head.pt, adapter/adapter_config.json, adapter/adapter_model.safetensors.
Reported common-test compute: BF16 Qwen backbone with FP32 candidate scoring.
Obtain the official Qwen backbone and tokenizer named in the configuration separately.
Inference scores the full state, question and complete candidate menu.

The supplied qwen_decision.py exposes the candidate-scoring inference interface. Supply the local package and the official backbone and tokenizer named in decision_config.json.

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.