Update project background and model performance

Browse files

Files changed (3) hide show

Model_Card_EN.md +54 -40
Model_Card_ZH.md +54 -40
README.md +54 -40

Model_Card_EN.md CHANGED Viewed

@@ -1,40 +1,54 @@
----
-language:
-- en
-license: other
-license_name: s2-mas-controlled-research-v1.2
-license_link: https://huggingface.co/ya9919/fin-laya-6-rounds/blob/main/LICENSE
-base_model: convaiinnovations/laya
-base_model_relation: finetune
-tags:
-- s2-mas
-- financial-one-shot
-- decision-model
----
-# Fin-Laya 6-rounds
-A financial one-shot decision model for S²-MAS. Inputs consist of a household state, a question and its complete candidate alternatives. Targets are observed responses from real-survey data. The model has approximately 421M parameters.
-Fin-Laya uses **2,925,263 post-training examples per language**, the largest disclosed domain post-training set among the financial decision models compared in the paper. The full-data MLX release achieves **74.5544% accuracy and 38.3746% question-macro F1**, exceeding Jev accuracy by **9.2960 percentage points** on the shared test.
-## Model files and inference
-This repository contains model weights, the tokenizer and the inference configuration. The native runtime is Laya / PyTorch with BF16 weights. `predict_example.py` uses an independently synthetic example.
-## Paper evaluation
-Evaluation: Native BF16.
-| Test | Records | Accuracy | Question-macro F1 |
-|---|---:|---:|---:|
-| Shared | 240,458 | 74.5586% | 38.3986% |
-| Full | 418,597 | 67.2728% | 27.8355% |
-The shared test contains 119 question versions. Question-macro F1 averages candidate-class F1 within each question version and then weights versions equally.
-## License and manuscript
-The project model contribution is offered under the [Controlled Model Research License v1.2](LICENSE). 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](https://github.com/KataDavidXD/s2-mas/blob/main/paper/S2_MAS_Paper.pdf) · [Project](https://github.com/KataDavidXD/s2-mas)

+---
+language:
+- en
+license: other
+license_name: s2-mas-controlled-research-v1.2
+license_link: https://huggingface.co/ya9919/fin-laya-6-rounds/blob/main/LICENSE
+base_model: convaiinnovations/laya
+base_model_relation: finetune
+tags:
+- s2-mas
+- financial-one-shot
+- decision-model
+---
+# Fin-Laya 6-rounds
+## 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. Fin-Laya 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 **native English-base Fin-Laya 6-round checkpoint**, with approximately 421M parameters. It continues the [3-round financial checkpoint](https://huggingface.co/ya9919/fin-laya-3-rounds) with SFT and RLCD over **2,925,263 training examples per language**, adapting Laya's ModernBERT-large encoder and option-scoring head to 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 [MLX runtime](https://huggingface.co/ya9919/fin-laya-mlx-6-rounds) deploys the same trained checkpoint family in FP16 and is the format used in the manuscript's English purchase-funding applications.
+## Paper results for this checkpoint
+| Test | Input language | Backend / precision | QA denominator | Question versions | Accuracy | Question-macro F1 |
+|---|---|---|---:|---:|---:|---:|
+| Shared financial questions | English | PyTorch BF16 | 240,458 | 119 | 74.5586% | 38.40% |
+| Expanded test | English | PyTorch BF16 | 418,597 | 862 | 67.2728% | 27.8355% |
+Accuracy weights QA records equally. Question-macro F1 averages F1 over the complete candidate classes within each question version, then weights question versions equally.
+Native shared-test F1 is shown at the precision displayed in Figure 5; Methods M6 reports the expanded test. The same checkpoint family in MLX FP16 scores 74.5544% accuracy and 38.3746% F1 on the shared test; see the [MLX card](https://huggingface.co/ya9919/fin-laya-mlx-6-rounds) and Methods M5.
+## Deployment timing
+The manuscript's single-request timing uses the [bilingual Fin-Laya checkpoint](https://huggingface.co/ya9919/fin-laya-multilingual-mlx) 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. Native Laya / PyTorch inference uses BF16. `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](LICENSE). 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](https://github.com/KataDavidXD/s2-mas/blob/main/paper/S2_MAS_Paper.pdf) · [Project](https://github.com/KataDavidXD/s2-mas)
+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.

Model_Card_ZH.md CHANGED Viewed

@@ -1,40 +1,54 @@
----
-language:
-- en
-license: other
-license_name: s2-mas-controlled-research-v1.2
-license_link: https://huggingface.co/ya9919/fin-laya-6-rounds/blob/main/LICENSE
-base_model: convaiinnovations/laya
-base_model_relation: finetune
-tags:
-- s2-mas
-- financial-one-shot
-- decision-model
----
-# Fin-Laya 6-rounds
-用于 S²-MAS 的金融一次性决策模型。输入为家庭状态、问题及完整候选选项，学习目标为真实调查中的已观察回答。模型约有421M参数。
-完整数据后训练集每语言含2,925,263个样本，是论文所比较金融决策模型中披露规模最大的领域后训练集。完整数据 MLX 版本在共享测试集上准确率为74.5544%，question-macro F1为38.3746%，准确率比Jev高9.2960个百分点。
-## 模型文件与推理
-本仓库包含模型权重、tokenizer和必要推理配置。使用原生 Laya / PyTorch 的 BF16 格式。 `predict_example.py` 使用独立合成的示例输入。
-## 论文结果
-口径：Native BF16。
-| 测试集 | 记录数 | Accuracy | Question-macro F1 |
-|---|---:|---:|---:|
-| 共享测试 | 240,458 | 74.5586% | 38.3986% |
-| 完整测试 | 418,597 | 67.2728% | 27.8355% |
-共享测试包含119个问题版本。Question-macro F1先在每个问题内平均候选类别F1，再对问题版本等权平均。
-## 许可与论文
-项目模型贡献适用[受限科研许可v1.2](LICENSE)。完整许可及保留的上游Apache-2.0权利见LICENSE。训练与评估记录不随模型发布。
-[S²-MAS论文](https://github.com/KataDavidXD/s2-mas/blob/main/paper/S2_MAS_Paper.pdf) · [项目](https://github.com/KataDavidXD/s2-mas)

+---
+language:
+- en
+license: other
+license_name: s2-mas-controlled-research-v1.2
+license_link: https://huggingface.co/ya9919/fin-laya-6-rounds/blob/main/LICENSE
+base_model: convaiinnovations/laya
+base_model_relation: finetune
+tags:
+- s2-mas
+- financial-one-shot
+- decision-model
+---
+# Fin-Laya 6-rounds
+## 项目背景
+S²-MAS 是面向社会规模的多智能体仿真框架，用于研究异质家庭在条件变化下的决策如何汇总为人口层面的结果。它将**5 亿条构造家庭 persona 资源**、可复用场景、可训练决策模型与持续更新的世界状态连接起来。Fin-Laya 在其中提供可复用的决策策略：根据家庭状态、问题和完整候选选项进行评分，使同一决策接口能够服务不同任务。
+论文中的配对条件应用实际评估了 **1,200 万个独立合成家庭 agent、共 2,400 万次选择**；24 个月扩展则跟踪 10,000 户家庭在收入、必要支出与银行 offer 变化下的状态。应用结果说明家庭资源如何改变消费响应：持续降收消耗现金储备并推迟购置，持续增收则同时支持现金储备与购置；偿付反馈进一步影响后续 offer，从而将个体选择接到消费、现金与支付的时间路径。
+## 本仓库的模型角色与训练
+本仓库提供**原生英语基座 Fin-Laya 6-round checkpoint**，约 421M 参数。它在 [3-round 金融 checkpoint](https://huggingface.co/ya9919/fin-laya-3-rounds)基础上，以**每语言 2,925,263 条训练样本**继续进行 SFT 与 RLCD，使 Laya 的 ModernBERT-large encoder 和选项评分 head 适配来自 real-survey data 的扩展监督。在论文比较且披露训练规模的金融决策模型中，这是最大的领域后训练样本集。
+[MLX 运行版本](https://huggingface.co/ya9919/fin-laya-mlx-6-rounds)以 FP16 部署同一训练 checkpoint 家族，也是论文英语购置融资应用使用的格式。
+## 本版本的论文结果
+| 测试口径 | 输入语言 | Backend / 精度 | QA 分母 | 问题版本 | Accuracy | Question-macro F1 |
+|---|---|---|---:|---:|---:|---:|
+| 共享金融问题 | 英语 | PyTorch BF16 | 240,458 | 119 | 74.5586% | 38.40% |
+| 扩展测试 | 英语 | PyTorch BF16 | 418,597 | 862 | 67.2728% | 27.8355% |
+Accuracy 对 QA 记录等权计分；question-macro F1 先在每个问题版本内平均完整候选类别的 F1，再对问题版本等权平均。
+原生共享测试的 F1 按 Figure 5 显示为 38.40%；扩展测试见 Methods M6。相同 checkpoint 家族的 MLX FP16 共享测试结果为 74.5544% accuracy、38.3746% F1，见 [MLX 版本卡片](https://huggingface.co/ya9919/fin-laya-mlx-6-rounds)及 Methods M5。
+## 部署速度
+论文的单请求速度测量使用[双语 Fin-Laya checkpoint](https://huggingface.co/ya9919/fin-laya-multilingual-mlx)，条件为一张 NVIDIA A100 PCIe 40GB、FP16、batch size 与并发均为 1、三个固定中文案例。对应延迟、吞吐量与 Jev 比较见该版本卡片及 Methods M9。
+## 模型文件与推理
+本仓库包含模型权重、tokenizer和必要推理配置。原生 Laya / PyTorch 推理使用 BF16。 `predict_example.py` 使用独立合成的示例输入。
+## 许可与论文
+项目模型贡献适用[受限科研许可v1.2](LICENSE)。完整许可及保留的上游Apache-2.0权利见LICENSE。训练与评估记录不随模型发布。
+[S²-MAS论文](https://github.com/KataDavidXD/s2-mas/blob/main/paper/S2_MAS_Paper.pdf) · [项目](https://github.com/KataDavidXD/s2-mas)
+公开项目提供科学制图、决策模型训练与评测，以及独立合成配套数据生成脚本；GENESE／MAS 引擎实现保持私有。

README.md CHANGED Viewed

@@ -1,40 +1,54 @@
----
-language:
-- en
-license: other
-license_name: s2-mas-controlled-research-v1.2
-license_link: https://huggingface.co/ya9919/fin-laya-6-rounds/blob/main/LICENSE
-base_model: convaiinnovations/laya
-base_model_relation: finetune
-tags:
-- s2-mas
-- financial-one-shot
-- decision-model
----
-# Fin-Laya 6-rounds
-A financial one-shot decision model for S²-MAS. Inputs consist of a household state, a question and its complete candidate alternatives. Targets are observed responses from real-survey data. The model has approximately 421M parameters.
-Fin-Laya uses **2,925,263 post-training examples per language**, the largest disclosed domain post-training set among the financial decision models compared in the paper. The full-data MLX release achieves **74.5544% accuracy and 38.3746% question-macro F1**, exceeding Jev accuracy by **9.2960 percentage points** on the shared test.
-## Model files and inference
-This repository contains model weights, the tokenizer and the inference configuration. The native runtime is Laya / PyTorch with BF16 weights. `predict_example.py` uses an independently synthetic example.
-## Paper evaluation
-Evaluation: Native BF16.
-| Test | Records | Accuracy | Question-macro F1 |
-|---|---:|---:|---:|
-| Shared | 240,458 | 74.5586% | 38.3986% |
-| Full | 418,597 | 67.2728% | 27.8355% |
-The shared test contains 119 question versions. Question-macro F1 averages candidate-class F1 within each question version and then weights versions equally.
-## License and manuscript
-The project model contribution is offered under the [Controlled Model Research License v1.2](LICENSE). 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](https://github.com/KataDavidXD/s2-mas/blob/main/paper/S2_MAS_Paper.pdf) · [Project](https://github.com/KataDavidXD/s2-mas)

+---
+language:
+- en
+license: other
+license_name: s2-mas-controlled-research-v1.2
+license_link: https://huggingface.co/ya9919/fin-laya-6-rounds/blob/main/LICENSE
+base_model: convaiinnovations/laya
+base_model_relation: finetune
+tags:
+- s2-mas
+- financial-one-shot
+- decision-model
+---
+# Fin-Laya 6-rounds
+## 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. Fin-Laya 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 **native English-base Fin-Laya 6-round checkpoint**, with approximately 421M parameters. It continues the [3-round financial checkpoint](https://huggingface.co/ya9919/fin-laya-3-rounds) with SFT and RLCD over **2,925,263 training examples per language**, adapting Laya's ModernBERT-large encoder and option-scoring head to 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 [MLX runtime](https://huggingface.co/ya9919/fin-laya-mlx-6-rounds) deploys the same trained checkpoint family in FP16 and is the format used in the manuscript's English purchase-funding applications.
+## Paper results for this checkpoint
+| Test | Input language | Backend / precision | QA denominator | Question versions | Accuracy | Question-macro F1 |
+|---|---|---|---:|---:|---:|---:|
+| Shared financial questions | English | PyTorch BF16 | 240,458 | 119 | 74.5586% | 38.40% |
+| Expanded test | English | PyTorch BF16 | 418,597 | 862 | 67.2728% | 27.8355% |
+Accuracy weights QA records equally. Question-macro F1 averages F1 over the complete candidate classes within each question version, then weights question versions equally.
+Native shared-test F1 is shown at the precision displayed in Figure 5; Methods M6 reports the expanded test. The same checkpoint family in MLX FP16 scores 74.5544% accuracy and 38.3746% F1 on the shared test; see the [MLX card](https://huggingface.co/ya9919/fin-laya-mlx-6-rounds) and Methods M5.
+## Deployment timing
+The manuscript's single-request timing uses the [bilingual Fin-Laya checkpoint](https://huggingface.co/ya9919/fin-laya-multilingual-mlx) 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. Native Laya / PyTorch inference uses BF16. `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](LICENSE). 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](https://github.com/KataDavidXD/s2-mas/blob/main/paper/S2_MAS_Paper.pdf) · [Project](https://github.com/KataDavidXD/s2-mas)
+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.