FinL-mlx · full six-round checkpoint (MLX FP16)

Model naming and experiment labels

FinL-mlx — full English-base six-round FinL checkpoint, MLX FP16 is the primary MLX runtime. The native Laya/PyTorch release is FinL. The separately evaluated bilingual deployment checkpoints are named FinL-bilingual and FinL-bilingual-mlx.

Project background

S²-MAS is a society-scale multi-agent simulation framework for studying how heterogeneous household decisions become population outcomes under changing conditions. It connects a constructed resource of 500 million household personas, reusable scenarios, trainable decision models and stateful world updates. FinL supplies a reusable decision policy: it scores the complete alternatives for a household state and question, allowing the same policy interface to serve different tasks.

The manuscript's paired-condition application evaluates 12 million distinct synthetic household agents and 24 million choices. A 24-month extension follows 10,000 households through changing income, expenses and bank offers. The applications show how household resources shape consumption responses: persistent income loss draws down reserves and delays purchases, while sustained income gains support both cash reserves and purchases. Repayment feedback affects later offers. These findings connect individual choices to consumption, cash and payment paths over time.

Repository role and training

This is the MLX FP16 runtime of the English-base FinL 6-round checkpoint, paired with the native release. The approximately 421M-parameter Laya encoder and option-scoring head continue the 3-round financial checkpoint through SFT and RLCD over 2,925,263 training examples per language, with expanded supervision from real-survey data. This is the largest disclosed domain post-training set among the financial decision models compared in the manuscript.

The manuscript's 12M-agent purchase-funding application and 24-month extension use this frozen checkpoint in MLX CUDA FP16.

Paper results for this checkpoint

Test	Input language	Backend / precision	QA denominator	Question versions	Accuracy	Question-macro F1
Shared financial questions	English	MLX CUDA FP16	240,458	119	74.5544%	38.3746%
Expanded test	English	MLX FP16	418,597	862	67.2726%	27.8955%

Accuracy weights QA records equally. Question-macro F1 averages F1 over the complete candidate classes within each question version, then weights question versions equally.

On the same English shared test, Jev 1.13 scores 65.2584% accuracy and 37.4793% F1; this release improves accuracy by 9.2960 percentage points. See Methods M5 and Table 4; Methods M6 reports the expanded test.

Deployment timing

The manuscript's single-request timing uses the FinL-bilingual checkpoint on one NVIDIA A100 PCIe 40GB in FP16, with batch size and concurrency one and three fixed Chinese cases. Its card and Methods M9 report latency, throughput and the Jev comparison.

Model files and inference

This repository contains model weights, the tokenizer and the inference configuration. The deployment format is Laya-MLX FP16. predict_example.py uses an independently generated synthetic example.

License and manuscript

The project model contribution is offered under the Controlled Model Research License v1.2. The full license and retained upstream Apache-2.0 rights are in LICENSE. Training and evaluation records are not distributed with the model.

S²-MAS manuscript · Project

The public project provides scientific plotting, decision-model training and evaluation, and independent synthetic companion-data generation scripts. The GENESE/MAS engine implementation remains private.