| ---
|
| language:
|
| - en
|
| - zh
|
| license: other
|
| license_name: s2-mas-controlled-research-v1.2
|
| license_link: https://huggingface.co/ya9919/fin-laya-multilingual-mlx/blob/main/LICENSE
|
| tags:
|
| - s2-mas
|
| - financial-one-shot
|
| - decision-model
|
| ---
|
|
|
| # FinL-bilingual-mlx 路 MLX CUDA FP16
|
|
|
| ## Model naming and experiment labels
|
|
|
| **FinL-bilingual-mlx** is the MLX FP16 runtime of the bilingual initial-supervision checkpoint, **FinL-bilingual**. The paper's A100 single-request timing protocol and 22.72脳 Jev shared-input-token-throughput result belong to this bilingual checkpoint pair. The main English-base six-round MLX runtime is [FinL-mlx](https://huggingface.co/ya9919/FinL-mlx).
|
|
|
| ## Project background
|
|
|
| S虏-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-bilingual 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 bilingual FinL-bilingual**, paired with the [native checkpoint](https://huggingface.co/ya9919/fin-laya-multilingual). The approximately 322M-parameter model adapts multilingual Laya's mmBERT encoder and option-scoring head to real-survey data, with two SFT stages and native RLCD on **1,721,623 training observations per language**. It supports English and Chinese inputs. Native and MLX packages provide runtime formats for the same bilingual trained checkpoint.
|
|
|
| ## Paper results for this checkpoint
|
|
|
| | Test | Input language | Backend / precision | QA denominator | Question versions | Accuracy | Question-macro F1 |
|
| |---|---|---|---:|---:|---:|---:|
|
| | Paired native checkpoint reference | Chinese | PyTorch BF16 | 240,458 | 119 | 74.52% | 37.99% |
|
|
|
| Accuracy weights QA records equally. Question-macro F1 averages F1 over the complete candidate classes within each question version, then weights question versions equally.
|
|
|
| The table is the paired trained checkpoint's native PyTorch BF16 reference; the MLX FP16 single-request timing is reported below. Figure 5 reports the Chinese shared test to two decimal places: accuracy improves by 21.61 percentage points and F1 by 10.60 points over the multilingual base.
|
|
|
| ## Deployment timing
|
|
|
| The **11 October 2026** deployment test uses this bilingual checkpoint on one **NVIDIA A100 PCIe 40GB**, with local models in **FP16**, **batch size and concurrency one**. Each of three fixed Chinese financial-choice cases receives three warm-ups and twenty timed repetitions, for sixty timed calls per arm.
|
|
|
| | Bilingual checkpoint deployment / reference | Mean latency (ms) | Choices/s | Shared input tokens/s |
|
| |---|---:|---:|---:|
|
| | FinL-bilingual 路 PyTorch FP16 | 22.06 | 45.33 | 15,247.34 |
|
| | FinL-bilingual-mlx 路 MLX CUDA FP16 | 9.10 | 109.88 | 36,955.47 |
|
| | Jev 1.13 路 API | 206.77 | 4.84 | 1,626.63 |
|
|
|
| MLX delivers **22.72 times Jev API's shared-input-token throughput** and **2.42 times the same checkpoint's PyTorch throughput**. Timing includes input preparation, tokenization, synchronized prediction and response formatting; loading is excluded and Jev includes its network round trip. Shared input tokens are counted with the mmBERT tokenizer on identical serialized requests. Jev's server hardware and precision are provider-managed. Methods M9 and Table S5 give the protocol and results.
|
|
|
| ## Inference
|
|
|
| Approximately 322M parameters. Runtime: Laya-MLX / MLX CUDA, FP16 deployment. Use `predict_example.py`, `example_request.json`, and caller-provided states with complete choice criteria. The packaged example is an independent synthetic input.
|
|
|
| Paired format: [native](https://huggingface.co/ya9919/fin-laya-multilingual).
|
|
|
| ## Paper and licensing
|
|
|
| Methods and results are described in the [S虏-MAS manuscript](https://github.com/KataDavidXD/s2-mas/blob/main/paper/S2_MAS_Paper.pdf).
|
|
|
| The project model contribution is governed by [LICENSE](LICENSE), which retains the original Apache-2.0 appendix and independent rights. 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.
|
|
|
| [Project](https://github.com/KataDavidXD/s2-mas)
|
|
|
