MedThoughts-8K

English|中文

This dataset is distilled from the full-scale DeepSeek-R1 (671B) in the medical domain. For more detailed information, please refer to our GitHub project MedR1.

1. Original Dataset

The data in this dataset is sourced from the US/train partition of MedQA (5 options).

2. Dataset Format

The keys in the dataset are explained as follows:

"question_id": The unique identifier for the question,
"question": The question itself,
"options": The options corresponding to the question (MedQA is a multiple-choice question dataset),
"answer_idx": The identifier for the correct option, such as A or B,
"ds_answer": The answer provided by DeepSeek-R1 (671B),
"ds_think": The reasoning process provided by DeepSeek-R1 (671B),
"source": The original dataset partition used, e.g., "MedQA/US/train" indicates that the `train` portion of the `US` (U.S. exam questions) is utilized.

3. Distillation Process

• 1. Download the MedQA dataset and format it into OpenAI message format to invoke the DeepSeek-R1 API;
  2. Use GPT-4o-mini to validate the correctness of the answers provided by DeepSeek-R1 (comparing the ground truth with DeepSeek-R1's responses);
  3. Filter out questions with incorrect answers, resulting in 7716 data entries to construct the MedThoughts-8K dataset;

4. Usage Recommendations

This dataset can be used for supervised fine-tuning (SFT) of existing language models to achieve models with reasoning capabilities similar to those of DeepSeek-R1.

You can use this script build_sft.py to convert this dataset into other sft formats, such as sharegpt, alpaca, query-response and messages into diverse file format (jsonl, json, csv, and xlsx).

example：

python build_sft.py --input_path xxx/xxx/MedThoughts-8K.jsonl --sft_format sharegpt --output_format json --output_dir xxx/xxx

5. Future Plans

We will soon make more distilled datasets in the medical field publicly available.

---

license: gpl-3.0