Fin-Laya-MLX

中文 | English

The two-source and four-source models are equally important, first-class releases, each with its own weights, evaluation and reproduction entry. This page describes four-source.

Release	Training sources	Download
two-source	CHFS, CGSS	rlcd-1
four-source	CHFS, CGSS, CFPS, CLDS	rlcd-2

The two-source release completed two SFT stages and one native RLCD stage. The four-source release continued with one SFT epoch and one RLCD epoch, using 2,925,263 English training records per current stage, six A100 GPUs and BF16. Single-choice examples are constructed from observed survey answers: CHFS at SWUFE, CGSS at RUC, CFPS at PKU and CLDS at SYSU.

Release	Test scope	Questions	Accuracy	Macro F1
two-source	CHFS + CGSS	240,458	74.5286%	38.0121%
two-source	CHFS + CGSS + CFPS + CLDS	418,597	56.2861%	19.2752%
four-source	CHFS + CGSS	240,458	74.5544%	38.3746%
four-source	CHFS + CGSS + CFPS + CLDS	418,597	67.2726%	27.8955%

Macro F1 averages over every candidate class within each supported question version, then equally over versions. The two test scopes are reported separately.

Release	Input tokens/s	Mean latency ms	Decisions/s
two-source	24,312	9.78	102.29
four-source	23,317	10.19	98.11

One A100 40GB, FP16, batch and concurrency one; the same three requests, three warmups and 20 timed calls per request. Tokens/s denotes input throughput using a shared tokenizer.

Approximately 0.42B parameters; an English financial decision component of the 1.4B Persona Social Model project. This version supports MLX CUDA.

versions.json · evaluation_report.json · evaluate.py · MAS Simulation

hf download ya9919/fin-laya-mlx --revision four-source --local-dir fin-laya-mlx-four-source
python evaluate.py --output evaluation_reproduced.json