---
license: apache-2.0
library_name: transformers
pipeline_tag: text-classification
tags: [ruhui, jev, laya, system-one, calibrated-decisions, rlcd, classification, routing, scoring, guardrails, moderation, reinforcement-learning, commercial-use]
---

# Ruhui · 如晦

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

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.

Architecture forked from [Laya](https://github.com/NandhaKishorM/laya) (Apache 2.0), with two key changes:

- **Chinese/multilingual backbone**: `mmBERT-base` (100+ languages) instead of English-only ModernBERT.
- **Bilingual soft-label fine-tuning**: 30+ domain datasets (intent / sentiment / safety / agent decision / tool-calling / …).

---

## Model Details

| Item | Value |
|---|---|
| Parameters | 322M (mmBERT-base + decision head) |
| Context length | 1024 |
| Head budget | 256 |
| Languages | Chinese, English, and 100+ |
| Training | RLCD (proper-scoring-rule policy gradient) + soft distillation + temperature calibration |

---

## Capabilities

Three decision primitives, evaluated in a single parallel forward pass:

| Primitive | Output | Use cases |
|---|---|---|
| `choice` | top label + full distribution + confidence | intent, routing, categorization |
| `score` | expected level on an ordinal rubric | urgency, frustration, severity |
| `noul` | calibrated P(true) | spam, phishing, jailbreak, churn risk |

Probabilities are trained with strictly proper scoring rules, so confidence is statistically meaningful and safe for confidence gating:

```python
if conf >= 0.85:
    route_automatically(dept)   # high confidence, no human in the loop
else:
    escalate_to_human(dept)     # low confidence, escalate
```

---

## Quick Start

Install the package first:

```bash
pip install ruhui
```

Then load the model and run typed decisions. Ruhui reads **Chinese and English (100+ languages)** in the same checkpoint — no separate English/multilingual models:

```python
import ruhui

agent = ruhui.load("anyforge/ruhui")

# Chinese input
result_zh = 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？"},
    },
)

# English input — same model, no switch
result_en = agent.predict(
    {"message": "I was charged twice, please refund me."},
    {
        "intent": {
            "type": "choice",
            "instructions": "What does the customer want?",
            "criteria": {"refund": "money back", "technical": "bug or outage", "billing": "invoice question"},
        },
        "churn_risk": {"type": "noul", "instructions": "Does the customer threaten to leave?"},
    },
)

print(result_zh["answers"])
print(result_en["answers"])
```

---

## Fine-Tuning

```bash
# 1. soft labels -> training items
python scripts/prepare_train_data.py --model_dir <base> --soft_dir <soft_labels> --out train_items.pt

# 2. train (RLCD + soft distillation + temperature calibration)
python scripts/train.py --model_dir <base> --train_items train_items.pt --output_dir output/ruhui --epochs 4

# 3. evaluate (Laya-aligned metrics)
python scripts/evaluate.py --model_dir output/ruhui --device cuda
```

See the [anyforge/ruhui](https://github.com/anyforge/ruhui) repository for details.

---

## License

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