Label the auxiliary FinL bilingual checkpoint

Model_Card_EN.md

@@ -11,17 +11,21 @@ tags:
 - decision-model
 ---
 
-# Fin-Laya Multilingual
+# FinL-bilingual · native PyTorch
+
+## Model naming and experiment labels
+
+**FinL-bilingual** is the bilingual initial-supervision checkpoint, with a multilingual encoder and English/Chinese support. Its MLX runtime is **FinL-bilingual-mlx**. The paper's A100 single-request timing protocol and 22.72× Jev shared-input-token-throughput result belong to this bilingual checkpoint pair. The main English-base six-round checkpoints are [FinL](https://huggingface.co/ya9919/FinL) and [FinL-mlx](https://huggingface.co/ya9919/FinL-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. 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.
+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-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 **native bilingual Fin-Laya checkpoint**, adapted from multilingual Laya's approximately 322M-parameter mmBERT encoder and option-scoring head. It supports English and Chinese inputs and uses **1,721,623 training observations per language** from real-survey data, with two SFT stages and native RLCD. It is the initial-supervision bilingual release, paired with the [MLX runtime](https://huggingface.co/ya9919/fin-laya-multilingual-mlx).
+This is the **native bilingual FinL-bilingual checkpoint**, adapted from multilingual Laya's approximately 322M-parameter mmBERT encoder and option-scoring head. It supports English and Chinese inputs and uses **1,721,623 training observations per language** from real-survey data, with two SFT stages and native RLCD. It is the initial-supervision bilingual release, paired with the [MLX runtime](https://huggingface.co/ya9919/fin-laya-multilingual-mlx).
 
 ## Paper results for this checkpoint
 
@@ -39,8 +43,8 @@ The **11 October 2026** deployment test uses this bilingual checkpoint on one **
 
 | Bilingual checkpoint deployment / reference | Mean latency (ms) | Choices/s | Shared input tokens/s |
 |---|---:|---:|---:|
-| Fin-Laya · PyTorch FP16 | 22.06 | 45.33 | 15,247.34 |
-| Fin-Laya · MLX CUDA FP16 | 9.10 | 109.88 | 36,955.47 |
+| FinL-bilingual · PyTorch FP16 | 22.06 | 45.33 | 15,247.34 |
+| FinL-bilingual-mlx · MLX CUDA FP16 | 9.10 | 109.88 | 36,955.47 |
 | Jev 1.13 · 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.

Model_Card_ZH.md

@@ -11,17 +11,21 @@ tags:
 - decision-model
 ---
 
-# Fin-Laya Multilingual
+# FinL-bilingual · native PyTorch
+
+## 模型命名与实验标签
+
+**FinL-bilingual** 是采用多语言 encoder、支持中英文的初始监督双语 checkpoint，其 MLX 格式名为 **FinL-bilingual-mlx**。论文的 A100 单请求速度协议及相对 Jev 的 22.72 倍共享输入 token 吞吐量结果属于这组双语 checkpoint。英语基座完整六轮主要版本为 [FinL](https://huggingface.co/ya9919/FinL) 与 [FinL-mlx](https://huggingface.co/ya9919/FinL-mlx)。
 
 ## 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。
+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。FinL-bilingual 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**原生双语 Fin-Laya checkpoint**，基于多语 Laya 的约 322M 参数 mmBERT encoder 与选项评分 head。模型支持英语和中文输入，以 real-survey data 的**每语言 1,721,623 条训练观测**完成两次 SFT 和一次原生 RLCD。它属于初始监督的双语发布版本，与 [MLX 运行版本](https://huggingface.co/ya9919/fin-laya-multilingual-mlx)配套。
+本仓库提供**原生双语 FinL-bilingual checkpoint**，基于多语 Laya 的约 322M 参数 mmBERT encoder 与选项评分 head。模型支持英语和中文输入，以 real-survey data 的**每语言 1,721,623 条训练观测**完成两次 SFT 和一次原生 RLCD。它属于初始监督的双语发布版本，与 [MLX 运行版本](https://huggingface.co/ya9919/fin-laya-multilingual-mlx)配套。
 
 ## 本版本的论文结果
 
@@ -39,8 +43,8 @@ Figure 5 按两位小数报告中文共享测试：相对多语基座，accuracy
 
 | 同一双语 checkpoint 的部署方式 / 参照 | 平均延迟 (ms) | Choices/s | 共享输入 tokens/s |
 |---|---:|---:|---:|
-| Fin-Laya · PyTorch FP16 | 22.06 | 45.33 | 15,247.34 |
-| Fin-Laya · MLX CUDA FP16 | 9.10 | 109.88 | 36,955.47 |
+| FinL-bilingual · PyTorch FP16 | 22.06 | 45.33 | 15,247.34 |
+| FinL-bilingual-mlx · MLX CUDA FP16 | 9.10 | 109.88 | 36,955.47 |
 | Jev 1.13 · API | 206.77 | 4.84 | 1,626.63 |
 
 MLX 的**共享输入 token 吞吐量**为 Jev API 的 **22.72 倍**，为同一 checkpoint PyTorch 的 **2.42 倍**。计时包含输入准备、tokenization、同步预测与响应格式化，模型加载不计入；Jev 包含网络往返。共享输入 token 使用相同序列化请求的 mmBERT tokenizer 计数。Jev 的服务端硬件与精度由服务方管理。协议与结果见 Methods M9、Table S5。

README.md

@@ -11,17 +11,21 @@ tags:
 - decision-model
 ---
 
-# Fin-Laya Multilingual
+# FinL-bilingual · native PyTorch
+
+## Model naming and experiment labels
+
+**FinL-bilingual** is the bilingual initial-supervision checkpoint, with a multilingual encoder and English/Chinese support. Its MLX runtime is **FinL-bilingual-mlx**. The paper's A100 single-request timing protocol and 22.72× Jev shared-input-token-throughput result belong to this bilingual checkpoint pair. The main English-base six-round checkpoints are [FinL](https://huggingface.co/ya9919/FinL) and [FinL-mlx](https://huggingface.co/ya9919/FinL-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. 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.
+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-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 **native bilingual Fin-Laya checkpoint**, adapted from multilingual Laya's approximately 322M-parameter mmBERT encoder and option-scoring head. It supports English and Chinese inputs and uses **1,721,623 training observations per language** from real-survey data, with two SFT stages and native RLCD. It is the initial-supervision bilingual release, paired with the [MLX runtime](https://huggingface.co/ya9919/fin-laya-multilingual-mlx).
+This is the **native bilingual FinL-bilingual checkpoint**, adapted from multilingual Laya's approximately 322M-parameter mmBERT encoder and option-scoring head. It supports English and Chinese inputs and uses **1,721,623 training observations per language** from real-survey data, with two SFT stages and native RLCD. It is the initial-supervision bilingual release, paired with the [MLX runtime](https://huggingface.co/ya9919/fin-laya-multilingual-mlx).
 
 ## Paper results for this checkpoint
 
@@ -39,8 +43,8 @@ The **11 October 2026** deployment test uses this bilingual checkpoint on one **
 
 | Bilingual checkpoint deployment / reference | Mean latency (ms) | Choices/s | Shared input tokens/s |
 |---|---:|---:|---:|
-| Fin-Laya · PyTorch FP16 | 22.06 | 45.33 | 15,247.34 |
-| Fin-Laya · MLX CUDA FP16 | 9.10 | 109.88 | 36,955.47 |
+| FinL-bilingual · PyTorch FP16 | 22.06 | 45.33 | 15,247.34 |
+| FinL-bilingual-mlx · MLX CUDA FP16 | 9.10 | 109.88 | 36,955.47 |
 | Jev 1.13 · 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.