Datasets:
File size: 2,690 Bytes
ab7addc 19418c0 ab7addc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 | ---
task_categories:
- question-answering
- text-generation
size_categories:
- 10K<n<100K
tags:
- medical
- reinforcement-learning
- onepo
- huatuogpt-3
configs:
- config_name: default
data_files:
- split: train
path: onepo_medical_20K.json
---
<div align="center">
<h1>OnePO-Medical-20K</h1>
<a href="https://openreview.net/pdf?id=M8eyUQldfx">📄 Paper</a> |
<a href="https://github.com/FreedomIntelligence/HuatuoGPT-3">💻 GitHub</a>
</div>
## ⚡ Introduction
**OnePO-Medical-20K** is the medical RL dataset released with **OnePO**, containing **20,338 medical tasks across multiple languages**.
- **One stage, no preceding SFT.** OnePO adapts pretrained models to medicine through a single reinforcement-learning stage.
- **Two complementary task types.** Multiple-choice questions provide verifiable answers. Open-ended conversations provide scoring rubrics.
- **Teacher guidance included.** Each task includes a HuatuoGPT-3 response for guidance during training.
## 📚 What's inside
| Task type | Samples | How responses are scored |
| --- | ---: | --- |
| **Multiple-choice** | 10,191 | Match the selected option against the correct answer. |
| **Open-ended** | 10,147 | Assess the response against weighted rubric criteria. |
**Data file:** `onepo_medical_20K.json` — a JSON array with one record per task.
| Fields | Description |
| --- | --- |
| `id`, `type` | Unique ID and task type: `multiple_choice` or `open_ended`. |
| `question`, `options`, `answer_idx` | Multiple-choice question, labeled options, and correct option label. |
| `prompt`, `rubrics` | Open-ended conversation and scoring criteria, each with `criterion` and signed `points`. |
| `teacher_response` | Teacher-generated response for training guidance, with reasoning in `<think>...</think>` followed by the final answer. |
Fields that do not apply to a task are `null`. Teacher responses can be replaced with outputs from another teacher model.
## 🚀 Usage
Load directly from Hugging Face:
```python
from datasets import load_dataset
dataset = load_dataset("FreedomIntelligence/OnePO-Medical-20K", split="train")
print(dataset[0])
```
To train with OnePO, follow the [repository setup](https://github.com/FreedomIntelligence/HuatuoGPT-3), then point `TRAIN_FILE` to the downloaded JSON:
```bash
TRAIN_FILE=/path/to/onepo_medical_20K.json bash OnePO.sh
```
## 📖 Citation
```bibtex
@inproceedings{chen2026onepo,
title={OnePO: Direct One-stage Policy Optimization for SFT-free Domain Adaptation},
author={Chen, Junying and Xie, Xinyuan and Li, Ziniu and Wang, Benyou},
booktitle={Proceedings of the 43rd International Conference on Machine Learning},
year={2026}
}
```
|