Ruhui · 如晦

A non-autoregressive System 1 decision engine for Chinese & multilingual text, with calibrated probabilities.

非自回归 System 1 决策引擎（中文/多语言），带校准概率。

Named after Du Ruhui (杜如晦, courtesy name Keming 克明) of the "Fang Mou Du Duan" (房谋杜断) pair — Fang Xuanling was the strategist, Du Ruhui the decisive judge. Ruhui inherits the "decisive" half: a fast System 1 decision maker that generates no text, has nothing to parse, and therefore cannot hallucinate.

命名取自「房谋杜断」的杜如晦（字克明），「晦」音近「hui」。房玄龄善谋、杜如晦善断——Ruhui 取「断」之意：System 1 快速决策，不生成文本、无可解析输出、因此无幻觉。

Architecture forked from Laya (Apache 2.0), with two key changes: 架构参照 Laya（Apache 2.0），在此基础上：

• Chinese/multilingual backbone: mmBERT-base (100+ languages) instead of English-only ModernBERT. 中文/多语言底座：mmBERT-base（100+ 语言），而非英语-only 的 ModernBERT。
• Bilingual soft-label fine-tuning: 30+ domain datasets (intent / sentiment / safety / agent decision / tool-calling / …). 中英双语软标签微调：覆盖 30+ 领域数据集（意图/情感/安全/智能体决策/工具调用等）。

Architecture · 架构

bidirectional encoder (mmBERT-base)
  → type injection (choice / score / noul)
  → 2-layer decision head
  → parallel scoring at [MASK] slots
  → softmax distribution (calibrated probabilities via strictly-proper-scoring-rule RLCD)

Three decision primitives · 三种决策原语：

Installation · 安装

pip install ruhui

Python 3.10+. Dependencies: torch, transformers, safetensors, huggingface_hub, numpy. 依赖：torch、transformers、safetensors、huggingface_hub、numpy。

Quick Start · 快速开始

import ruhui

# Load the fine-tuned ruhui checkpoint (from the hub, or a local directory)
# 加载微调后的 ruhui checkpoint（从仓库拉取，或本地目录路径）
agent = ruhui.load("anyforge/ruhui")

# The Router also accepts a local path (skips hub download):
# Router 同样支持本地路径（不走仓库下载）：
router = ruhui.Router(model_path="./pretrained/ruhui")
result = router.predict(state, questions)

# Answer multiple structured questions in a single forward pass
# 一次前向传播回答多个结构化问题
result = agent.predict(
    {"message": "我被重复扣款了，请退款"},
    {
        "intent": {
            "type": "choice",
            "instructions": "客户想做什么？",
            "criteria": {"refund": "退款", "technical": "技术问题", "billing": "账单咨询"},
        },
        "churn_risk": {
            "type": "noul",
            "instructions": "客户是否威胁要离开？",
        },
    },
)
print(result["answers"])

Fine-Tuning · 微调

Fine-tune ruhui on your own domain data (RLCD + soft distillation + temperature calibration). 在自己的领域数据上微调（RLCD + 软蒸馏 + 温度校准）。

1. Prepare soft-label data · 准备软标签数据 datas/soft_*.jsonl, each line: text + qtype + soft_label.probabilities. 每行含 text + qtype + soft_label.probabilities。

2. Build training items · 转训练 items：

   python3 scripts/prepare_train_data.py \
     --model_dir <base_model_dir> \
     --soft_dir <soft_label_dir> \
     --out <train_items.pt>
   

3. Train · 训练：

   python3 scripts/train.py \
     --model_dir <base_model_dir> \
     --train_items <train_items.pt> \
     --output_dir <output_dir> \
     --epochs 4
   

4. Inference · 推理：

   python3 scripts/predict.py --model_dir <output_dir> --text "..." \
     --question-type noul --instruction "..."
   

Repository Layout · 目录结构

ruhui/
  ruhui/            # package · 包（agent / router / common / presets / …）
  scripts/          # fine-tuning + inference scripts · 微调 + 推理脚本
    prepare_train_data.py
    train.py
    predict.py
  tests/
  pyproject.toml

Model Repositories · 模型仓库

• Hugging Face: anyforge/ruhui
• ModelScope: anyforge/ruhui

Single checkpoint: fine-tuned from mmBERT-base, bilingual (Chinese/English). Load it directly with ruhui.load("anyforge/ruhui"). 单一 checkpoint：基于 mmBERT-base 微调，中英双语。直接 ruhui.load("anyforge/ruhui") 加载。

License

Apache 2.0 (inherited from Laya). Developed by AnyForge. Apache 2.0（参照 Laya）。Developed by AnyForge。