🍎 Synthetic Multipersona Doctor Patient Conversations.

Author: Nisten Tahiraj

License: MIT

🧠 Generated by DeepSeek V3 running in full BF16.

🛠️ Done in a way that includes induced errors/obfuscations by the AI patients and friendly rebutals and corrected diagnosis from the AI doctors. This makes the dataset very useful as both training data and retrival systems for reducing hallucinations and increasing the diagnosis quality.

🐧 Conversations generated in the Following languages

_

English

Chinese

Japanese

Danish

German

French

_

More languages coming :) Follow our org lead by Doctor @JohnsonThomasMD for more updates, DeepSeek R1 generations and a new mobile opensource medical model are in the works too 🚀 .

The following disease list was used as seed for each synthetic convo: nisten/all-human-diseases

DISCLAIMER: These are not human conversations. These were not corrected by a human at all. These are all straight from the AI.

Before the data was generated the medical performance of the LLM was measured to be significantly higher than even Google's MedPalm 2.

Reference: MedPalm two scores no higher than 72% https://paperswithcode.com/sota/multiple-choice-question-answering-mcqa-on-21

Despite the driver issues, deepseek v3 instruct has stellar scores in medical benmarking, here running in fp8_w8a8 on 8x AMD Mi300x card the multimedqa bench. Little to no difference was observed in medical benchmarking in bfloat16 vs 8bit. However other tests showed some divergence: https://x.com/nisten/status/1874996106540503367

Yes, raw deepseek v3 with no special prompting scores 79% vs only 72% for the complicated CoT MedPalm2 API setup.

The newer DeepSeek R1 has not yet been tested.

Feel free to leave comments, concerns, and even contribute more data to open science.

Thank you https://www.vultr.com/ for sponsoring the compute.