ZombitX64/Medical-o1-Reasoning-SFT-Thai

Medical-GPT-Reasoning-Thai

Dataset Summary

This dataset contains medical Q&A data in JSON format, designed for fine-tuning AI models in medical reasoning and response generation.representing a medical question, complex chain-of-thought reasoning, and a concise response. All content is in Thai language.

The dataset is derived from a larger medical Q&A collection and has been processed to ensure JSON validity, with multi-line objects combined into single valid entries.

Supported Tasks and Leaderboards

• —Tasks: Question Answering, Text Generation, Medical Reasoning
• —Languages: Thai (th)

Languages

Thai (th) - all text is in Thai language.

Dataset Structure

Data Instances

Each instance is a JSON object with the following structure:

{
  "Question": "Medical question in Thai",
  "Complex_CoT": "Detailed chain-of-thought reasoning in Thai, including analysis, differential diagnosis, and management steps",
  "Response": "Concise medical response in Thai"
}

Data Fields

• —Question (string): The medical question asked by a patient or user.
• —Complex_CoT (string): Detailed reasoning process including analysis, differential diagnosis, and management recommendations.
• —Response (string): Concise final answer or response to the question.

Data Splits

The dataset contains a single split:

• —train: 16000 examples

Dataset Creation

Curation Rationale

This dataset was created to provide high-quality medical Q&A data for training AI models in Thai medical contexts, focusing on structured reasoning and accurate responses.

Source Data

• —Initial Data Collection and Normalization: Derived from Final-o1-sft-medical-15k.jsonl with extensive cleaning and validation.
• —Who are the source language producers?: Medical professionals and AI-generated content.
• —Annotations: Chain-of-thought reasoning was included to enhance model understanding.

Annotations

• —Annotation process: Manual and automated processing to ensure validity.
• —Who are the annotators?: Data processing scripts and human verification.

Personal and Sensitive Information

The dataset contains medical information but no personal identifiable information (PII).

Considerations for Using the Data

Social Impact of Dataset

This dataset can help improve AI systems for medical assistance in Thai-speaking regions, potentially improving healthcare access.

Discussion of Biases

• —The dataset is focused on Thai medical content and may not generalize to other languages or medical systems.
• —Content is generated and may contain biases from the original sources.

Additional Information

Dataset Curators

[ZombitX64/organization]

Licensing Information

cc-by-sa-4.0 License

Citation Information

@dataset{Medical-GPT-Reasoning-Thai,
  title={Medical-GPT-Reasoning-Thai Dataset},
  author={ZombitX64 , krittanut janutsaha, chanyut saengwichain},
  year={2025},
  url={https://huggingface.co/datasets/ZombitX64/o1-sft-medical}
}

Introduction

This dataset is used to fine-tune HuatuoGPT-o1, a medical LLM designed for advanced medical reasoning. This dataset is constructed using GPT-4o, which searches for solutions to verifiable medical problems and validates them through a medical verifier.

For details, see our paper and GitHub repository.

Citation

If you find our data useful, please consider citing our work!

@misc{chen2024huatuogpto1medicalcomplexreasoning,
      title={HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs}, 
      author={Junying Chen and Zhenyang Cai and Ke Ji and Xidong Wang and Wanlong Liu and Rongsheng Wang and Jianye Hou and Benyou Wang},
      year={2024},
      eprint={2412.18925},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.18925}, 
}

Contributions

Thanks to the contributors who helped process and validate the data.

Usage

Loading in Python

import json

with open('train.json', 'r', encoding='utf-8') as f:
    for line in f:
        data = json.loads(line.strip())
        print(data['Question'])
        print(data['Complex_CoT'])
        print(data['Response'])

Using with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset('json', data_files='train.json')
print(dataset['train'][0])

Fine-Tuning Example

This dataset can be used for supervised fine-tuning of language models like GPT or LLaMA for medical Q&A tasks.

from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import load_dataset

# Load dataset
dataset = load_dataset('json', data_files='train.json')

# Load model and tokenizer
model_name = "your-model-name"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Fine-tuning code here...

Notes

• —All text is in Thai language.
• —The dataset focuses on medical reasoning with structured chain-of-thought.
• —If you need the original line count (~17k), consider using alternative processing to split combined objects.
• —For any issues with specific lines, refer to the processing scripts in the repository.