CrossNow/Medical-Reasoning-SFT-Mega

---
license: apache-2.0
task_categories:
- text-generation
- question-answering
language:
- en
tags:
- medical
- reasoning
- healthcare
- clinical
- chain-of-thought
- thinking
- sft
- mega
- combined
size_categories:
- 1M<n<10M
---

# Medical-Reasoning-SFT-Mega

The ultimate medical reasoning dataset - combining 7 state-of-the-art AI models with fair distribution deduplication. 1.79 million unique samples with 3.78 billion tokens of medical chain-of-thought reasoning.

## Dataset Overview

| Metric | Value |
|--------|-------|
| **Total Samples** | 1,789,998 (after deduplication) |
| **Total Tokens** | ~3.78 Billion |
| **Content Tokens** | ~2.22 Billion |
| **Reasoning Tokens** | ~1.56 Billion |
| **Samples with Reasoning** | 1,789,764 (100.0%) |
| **Unique Questions** | 1,237,711 |
| **Shared Questions** | 552,287 |
| **Source Models** | 7 |
| **Language** | English |

## Source Models

This dataset combines reasoning from 7 leading AI models with fair representation:

| Model | Original | Unique | From Shared | Final | % |
|-------|----------|--------|-------------|-------|---|
| MiniMax-M2.1 | 204,773 | 76,892 | 42,126 | 119,018 | 6.6% |
| Baichuan-M3-235B | 124,520 | 30,589 | 43,070 | 73,659 | 4.1% |
| GPT-OSS-120B | 506,150 | 403,208 | 42,113 | 445,321 | 24.9% |
| Qwen3-Next-80B | 604,249 | 295,828 | 118,539 | 414,367 | 23.1% |
| GLM_4.5_Air | 225,179 | 151,278 | 42,117 | 193,395 | 10.8% |
| Nemotron-Nano-30B | 444,544 | 86,978 | 118,546 | 205,524 | 11.5% |
| Trinity-Mini | 810,284 | 192,938 | 145,776 | 338,714 | 18.9% |

**Total before deduplication:** 2,919,699 samples
**Total after deduplication:** 1,789,998 samples

## Deduplication Strategy: Fair Distribution

When the same medical question appears in multiple source datasets, we use **fair distribution** to ensure balanced representation:

1. **Unique questions** (1,237,711): Questions appearing in only one model are included directly
2. **Shared questions** (552,287): Questions appearing in multiple models are distributed fairly across all contributing models

For shared questions, we assign each question to the model that has been assigned the fewest shared questions so far. This ensures no single model dominates the duplicate pool, resulting in diverse reasoning styles across the dataset.

This differs from priority-based deduplication where one model would "win" all duplicates. With fair distribution, each model gets roughly equal share of the shared question pool.

## Schema

Each sample follows the conversational messages format with reasoning content:

```json
{
"messages": [
{
"role": "system",
"content": "You are a medical expert...",
"reasoning_content": null
},
{
"role": "user",
"content": "What are the symptoms of diabetes?",
"reasoning_content": null
},
{
"role": "assistant",
"content": "The main symptoms of diabetes include...",
"reasoning_content": "Let me think through this systematically..."
}
]
}
```

### Fields

| Field | Type | Description |
|-------|------|-------------|
| `messages` | list | Array of message objects in the conversation |
| `messages[].role` | string | Either "system", "user", or "assistant" |
| `messages[].content` | string | The main message content |
| `messages[].reasoning_content` | string or null | Chain-of-thought reasoning (assistant messages only) |

## Usage

### Loading with Datasets Library

```python
from datasets import load_dataset

dataset = load_dataset("OpenMed/Medical-Reasoning-SFT-Mega")
```

### Accessing Samples

```python
# Get a sample
sample = dataset['train'][0]

# Access messages
for msg in sample['messages']:
print(f"Role: {msg['role']}")
print(f"Content: {msg['content'][:100]}...")
if msg['reasoning_content']:
print(f"Reasoning: {msg['reasoning_content'][:100]}...")
```

### Filtering by Reasoning

```python
# Get samples with reasoning content
samples_with_reasoning = dataset['train'].filter(
lambda x: x['messages'][-1]['reasoning_content'] is not None
)
```

## Intended Use

This dataset is designed for:

- **Fine-tuning medical reasoning models**: Train LLMs with diverse reasoning styles from multiple state-of-the-art models
- **Chain-of-thought training**: Develop models that show detailed thinking processes
- **Medical QA systems**: Build robust question-answering systems for healthcare applications
- **Research**: Study and compare reasoning patterns across different AI architectures
- **Distillation**: Transfer capabilities from multiple large models to smaller ones

## Why Mega?

1. **Diversity**: 7 different model architectures provide varied reasoning approaches
2. **Fair Representation**: Fair distribution ensures balanced contributions from all models
3. **Scale**: 1.79M unique samples and 3.78B tokens for comprehensive training
4. **Coverage**: Spans clinical, diagnostic, pharmacological, and general medical knowledge

## Limitations and Considerations

- This dataset is generated by AI models and should not be used as a substitute for professional medical advice
- Responses may contain inaccuracies and should be validated by medical professionals
- Not intended for clinical decision-making without expert review
- The reasoning traces reflect model approaches, not necessarily optimal clinical reasoning

## Related Datasets

Individual model datasets are also available:

- [Medical-Reasoning-SFT-MiniMax-M2.1](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-MiniMax-M2.1)
- [Medical-Reasoning-SFT-Baichuan-M3-235B](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Baichuan-M3-235B)
- [Medical-Reasoning-SFT-GPT-OSS-120B-V2](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-GPT-OSS-120B-V2)
- [Medical-Reasoning-SFT-Qwen3-Next-80B](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Qwen3-Next-80B)
- [Medical-Reasoning-SFT-GLM_4.5_Air](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-GLM_4.5_Air)
- [Medical-Reasoning-SFT-Nemotron-Nano-30B](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Nemotron-Nano-30B)
- [Medical-Reasoning-SFT-Trinity-Mini](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Trinity-Mini)

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{medical_reasoning_sft_mega,
title={Medical-Reasoning-SFT-Mega},
author={OpenMed},
year={2025},
publisher={Hugging Face},
url={https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Mega}
}
```

## License

Apache 2.0

---
许可证：Apache-2.0
任务类别：
- 文本生成
- 问答
语言：
- 英语
标签：
- 医疗
- 推理
- 医疗保健
- 临床
- 思维链（Chain-of-Thought）
- 思考
- 监督微调（Supervised Fine-Tuning，SFT）
- 大型（Mega）
- 组合
样本规模类别：
100万<样本数<1000万
---

# 《医疗推理-监督微调-大型合集》（Medical-Reasoning-SFT-Mega）

本数据集为顶级医疗推理数据集，整合了7个当前最优的人工智能模型，并采用公平分布去重策略。数据集包含1,789,998条唯一样本，总令牌数达37.8亿，涵盖医疗领域思维链推理内容。

## 数据集概览
| 指标 | 数值 |
|--------|-------|
| **总样本数** | 1,789,998（去重后） |
| **总令牌数** | ~37.8亿 |
| **内容令牌数** | ~22.2亿 |
| **推理令牌数** | ~15.6亿 |
| **含推理内容的样本数** | 1,789,764（占比100.0%） |
| **唯一问题数** | 1,237,711 |
| **共享问题数** | 552,287 |
| **源模型数量** | 7 |
| **语言** | 英语 |

## 源模型
本数据集整合了7款主流人工智能模型的推理结果，各模型贡献均衡：
| 模型名称 | 原始样本数 | 唯一样本数 | 共享样本数 | 最终样本数 | 占比 |
|-------|----------|--------|-------------|-------|---|
| MiniMax-M2.1 | 204,773 | 76,892 | 42,126 | 119,018 | 6.6% |
| Baichuan-M3-235B | 124,520 | 30,589 | 43,070 | 73,659 | 4.1% |
| GPT-OSS-120B | 506,150 | 403,208 | 42,113 | 445,321 | 24.9% |
| Qwen3-Next-80B | 604,249 | 295,828 | 118,539 | 414,367 | 23.1% |
| GLM_4.5_Air | 225,179 | 151,278 | 42,117 | 193,395 | 10.8% |
| Nemotron-Nano-30B | 444,544 | 86,978 | 118,546 | 205,524 | 11.5% |
| Trinity-Mini | 810,284 | 192,938 | 145,776 | 338,714 | 18.9% |

**去重前总样本数：** 2,919,699
**去重后总样本数：** 1,789,998

## 去重策略：公平分布法
当同一医疗问题出现在多个源数据集中时，本数据集采用**公平分布**策略以确保各模型贡献均衡：
1. **唯一问题（1,237,711条）**：仅在单个模型中出现的问题将被直接收录
2. **共享问题（552,287条）**：在多个模型中出现的问题将被公平分配至所有贡献模型

对于共享问题，我们会将其分配给当前已分配共享问题数量最少的模型。该策略可避免单一模型主导重复样本池，确保数据集内推理风格的多样性。
此策略与基于优先级的去重不同——后者会让某一个模型“独占”所有重复样本。通过公平分布，每个模型在共享问题池中均可获得大致均等的份额。

## 数据格式
每条样本遵循带推理内容的会话消息格式：
json
{
"messages": [
{
"role": "system",
"content": "You are a medical expert...",
"reasoning_content": null
},
{
"role": "user",
"content": "What are the symptoms of diabetes?",
"reasoning_content": null
},
{
"role": "assistant",
"content": "The main symptoms of diabetes include...",
"reasoning_content": "Let me think through this systematically..."
}
]
}

### 字段说明
| 字段 | 类型 | 描述 |
|-------|------|-------------|
| `messages` | 列表 | 会话消息对象数组 |
| `messages[].role` | 字符串 | 角色，仅支持"system"、"user"或"assistant" |
| `messages[].content` | 字符串 | 消息主体内容 |
| `messages[].reasoning_content` | 字符串或空 | 思维链推理内容（仅助手消息包含此字段） |

## 使用方法
### 通过Datasets库加载
python
from datasets import load_dataset

dataset = load_dataset("OpenMed/Medical-Reasoning-SFT-Mega")

### 访问样本
python
# 获取单条样本
sample = dataset['train'][0]

# 遍历会话消息
for msg in sample['messages']:
print(f"角色：{msg['role']}")
print(f"内容：{msg['content'][:100]}...")
if msg['reasoning_content']:
print(f"推理内容：{msg['reasoning_content'][:100]}...")

### 按推理内容筛选
python
# 获取包含推理内容的样本
samples_with_reasoning = dataset['train'].filter(
lambda x: x['messages'][-1]['reasoning_content'] is not None
)

## 预期用途
本数据集适用于以下场景：
- **医疗推理模型微调**：利用多款当前最优模型的多样化推理风格对大语言模型（Large Language Model，LLM）进行微调
- **思维链训练**：开发可展示详细思考过程的模型
- **医疗问答系统**：构建适用于医疗健康场景的稳健问答系统
- **研究工作**：研究并对比不同人工智能架构的推理模式
- **模型蒸馏**：将多大型模型的能力迁移至小型模型

## 为何命名为“大型合集”？
1. **多样性**：7种不同的模型架构提供了多样化的推理思路
2. **公平贡献**：公平分布策略确保各模型的贡献保持均衡
3. **规模性**：179万条唯一样本与37.8亿令牌，可支撑全面的模型训练
4. **覆盖范围**：涵盖临床、诊断、药理及通用医疗知识领域

## 局限性与注意事项
- 本数据集由人工智能模型生成，不可替代专业医疗建议
- 模型输出可能存在不准确之处，需由医疗专业人员进行验证
- 未经专家审核，不得将其用于临床决策
- 推理轨迹仅反映模型自身的推理思路，未必等同于最优的临床推理逻辑

## 相关数据集
各单模型数据集亦可单独获取：
- [Medical-Reasoning-SFT-MiniMax-M2.1](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-MiniMax-M2.1)
- [Medical-Reasoning-SFT-Baichuan-M3-235B](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Baichuan-M3-235B)
- [Medical-Reasoning-SFT-GPT-OSS-120B-V2](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-GPT-OSS-120B-V2)
- [Medical-Reasoning-SFT-Qwen3-Next-80B](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Qwen3-Next-80B)
- [Medical-Reasoning-SFT-GLM_4.5_Air](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-GLM_4.5_Air)
- [Medical-Reasoning-SFT-Nemotron-Nano-30B](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Nemotron-Nano-30B)
- [Medical-Reasoning-SFT-Trinity-Mini](https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Trinity-Mini)

## 引用格式
若您使用本数据集，请引用以下内容：
bibtex
@dataset{medical_reasoning_sft_mega,
title={Medical-Reasoning-SFT-Mega},
author={OpenMed},
year={2025},
publisher={Hugging Face},
url={https://huggingface.co/datasets/OpenMed/Medical-Reasoning-SFT-Mega}
}

## 许可证
Apache 2.0