Fin-Laya-Multilingual-MLX — 3-rounds

Training release: 3-rounds (SFT → SFT → RLCD).

Fin-Laya Models collection

中文 | English

English

Financial decision model

Fin-Laya-Multilingual adapts Laya multilingual to financial behavior using observed survey answers. It supports structured choice tasks in Chinese and English. The model is a financial behavior component of the 1.4B Persona Social Model project, covering household consumption, borrowing, saving, and repayment choices.

Current version: MLX / MLX CUDA

Data and deployment

The dataset contains 2,455,152 real survey question–answer records across 119 question versions for training and evaluation. Sources: China Household Finance Survey (CHFS) and Chinese General Social Survey (CGSS).

With approximately 0.32B parameters and 644 MB of FP16 weights, the model is suitable for deployment on consumer GPUs. PyTorch and MLX versions are available. Fin-Laya-Multilingual-MLX supports MLX CUDA and has been verified on one NVIDIA A100.

Accuracy and performance

Model	Accuracy	Macro F1
Original Laya multilingual	52.91%	27.39%
Fin-Laya-Multilingual · After two SFT stages	74.49%	37.58%
Fin-Laya-Multilingual · Two SFT stages + one RLCD stage	74.52%	37.99%

Matched held-out test set: 240,458 question–answer records. Macro F1 is computed within each question and then averaged across the 119 question versions.

Single-GPU inference speed

Model and runtime	Mean latency (ms)	Decisions/s	Input tokens/s
Original Laya multilingual · PyTorch	21.84	45.78	15,397
Fin-Laya-Multilingual · PyTorch	22.06	45.33	15,247
Fin-Laya-Multilingual · MLX CUDA	9.10	109.88	36,955
Jev 1.13.0 · API	206.77	4.84	1,627

Laya runs on the same NVIDIA A100 40GB with one request at a time. The benchmark uses three identical Chinese financial cases, with three warmups and 20 timed calls per case. Jev is an API reference; its latency includes network and service time.

tokens/s measures input processing using a shared tokenizer. Laya returns structured choices.

Training: two SFT stages, one RLCD stage

1. SFT stage 1: start from Laya multilingual and learn observed respondent answers across 21 question versions.
2. SFT stage 2: expand to 119 question versions and continue supervised financial adaptation.
3. RLCD stage 3: start from the second SFT checkpoint and continue training with Laya’s native RLCD.

Training uses six A100 GPUs and BF16. The released checkpoint follows SFT → SFT → RLCD.

Model versions and reports

• PyTorch version
• MLX version (MLX CUDA supported)
• Training and evaluation report · Inference speed report
• GitHub

Reproduce the evaluation

Frozen test questions and observed answers: 240,458 records per language (zh). evaluation_report.json contains overall and per-question scores.

F1 first averages class F1 across every candidate class within each supported question version, with zero denominators assigned zero. It then averages equally across question versions with test records. Accuracy is total correct predictions divided by test records.

python evaluate.py --output evaluation_reproduced.json

See versions.json for historical versions, for example python evaluate.py --version rlcd-1.

The existing result for this older MLX revision is a 378-case conversion agreement check. Its full-test PyTorch reference scores are identified separately in the report. The command above computes this MLX revision’s own full-test scores.