---
title: RiverONE
canonical_url: "https://www.modelscope.cn/models/ThewakeRiverYtzLab/RiverONE"
md_url: "https://www.modelscope.cn/models/ThewakeRiverYtzLab/RiverONE.md"
repository: ThewakeRiverYtzLab/RiverONE
chinese_name: RiverONE
last_updated: 2026-08-31
model_type:
  - riverone_qc
architectures:
  - RiverOneQCModel
parameters: 3.0B
tensor_type:
  - BF16
  - I32
  - F32
library_name:
  - pytorch
  - transformer
  - safetensors
downloads: 27
stars: 0
---

# RiverONE

> RiverONE - ThewakeRiverYtzLab 在 ModelScope 开源的模型。QCalEval ZeroShot 评测结果

- **Repository**: ThewakeRiverYtzLab/RiverONE
- **Parameters**: 3.0B
- **Downloads**: 27
- **Stars**: 0
- **Last updated**: 2026-08-31

Source: https://www.modelscope.cn/models/ThewakeRiverYtzLab/RiverONE

---

# RiverONE

## QCalEval ZeroShot 评测结果

评测集: RiverOne_ZeroShot_v1.2_Test (243样本)

| 题目 | 得分 | 评测方式 | 说明 |
|---|---|---|---|
| Q1 图表理解 | **63.63%** | LLM Judge (GPT-5) | 程序化字段(50%) + key_points(50%) |
| Q2 图表分类 | **86.83%** | 精确匹配 | 图表类型分类 (211/243) |
| Q3 关键点提取 | **46.85%** | LLM Judge (GPT-5) | key_points 逐项打分 |
| Q4 评估判断 | **96.30%** | 精确匹配 | Assessment 分类 (234/243) |
| Q5 字段提取 | **83.55%** | 容差规则 | 字段容差打分 |
| Q6 状态分类 | **85.60%** | 精确匹配 | Status 分类 (208/243) |
| **平均** | **77.13%** | | |

## 快速使用

### 环境要求

```bash
pip install transformers>=4.45 torch>=2.0 aqlm
```

### 加载模型

```python
import torch
from transformers import AutoTokenizer, AutoConfig
from modeling_riverone_qc import RiverOneQCModel

# 加载模型配置
config = AutoConfig.from_pretrained(
    "path/to/this/model",
    trust_remote_code=True
)

# 加载模型
model = RiverOneQCModel.from_pretrained(
    "path/to/this/model",
    config=config,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

# 加载 tokenizer
tokenizer = AutoTokenizer.from_pretrained(
    "path/to/this/model",
    trust_remote_code=True
)
```

### 推理示例

```python
from PIL import Image

# 加载图像
image = Image.open("chart.png").convert("RGB")

# 构建对话
messages = [
    {"role": "user", "content": "<image>\n请描述这张图表。"}
]

# 推理
response = model.chat(
    tokenizer=tokenizer,
    pixel_values=image,
    messages=messages,
    max_new_tokens=512
)
print(response)
```

## 目录结构

```
.
├── config.json                    # 模型配置
├── quant_config.json              # AQLM 量化配置
├── generation_config.json         # 生成配置
├── model.safetensors              # 模型权重 (6.9 GB)
├── model.safetensors.index.json   # 权重索引
├── tokenizer.json                 # Tokenizer
├── tokenizer_config.json          # Tokenizer 配置
├── added_tokens.json              # 额外 token
├── special_tokens_map.json        # 特殊 token 映射
├── chat_template.jinja            # 对话模板
├── configuration_riverone_qc.py   # 模型配置类
├── modeling_riverone_qc.py        # 模型实现
├── modeling_ising_vit.py          # ViT 实现
├── processor_config.json          # 处理器配置
├── preprocessor_config.json       # 预处理器配置
├── video_preprocessor_config.json # 视频预处理器配置
└── README.md                      # 本文件
```

## 引用

本模型基于以下工作：

- **Qwen3**: Qwen Technical Report
- **AQLM**: Extreme Compression of Large Language Models via Additive Quantization
- **InternVL**: InternVL: Scaling up Vision Foundation Models
- **IsingViT**: Quantum-inspired Vision Transformer