🧠 Medical-Reasoning-CoT-MegaMix

A curated mixture of medical reasoning datasets from multiple teacher models, formatted for fine‑tuning Llama‑3.1‑based models to perform step‑by‑step clinical reasoning. The dataset combines explicit chain‑of‑thought traces with general instruction data to preserve broad conversational abilities.

📊 Dataset Composition

Source	Target Samples	Reasoning Type
OpenMed/Medical-Reasoning-SFT-Mega	20,000	Mixed teacher reasoning (implicit in messages)
OpenMed/Medical-Reasoning-SFT-Qwen3-Next-80B	15,000	High‑IQ teacher reasoning (implicit)
OpenMed/Medical-Reasoning-SFT-GPT-OSS-120B-V2	15,000	GPT‑OSS teacher reasoning (implicit)
OpenMed/Medical-Reasoning-SFT-Baichuan-M3-235B	15,000	Baichuan teacher reasoning (implicit)
Intelligent-Internet/II-Medical-Reasoning-SFT	15,000	Multi‑turn reasoning (implicit in responses)
FreedomIntelligence/medical-o1-reasoning-SFT	20,000	Explicit chain‑of‑thought (ideal for CoT training)
teknium/OpenHermes-2.5	35,000	General instruction‑following data
Total	~135,000	Balanced mix (~74% medical / 26% general)

All medical sources are filtered to include only examples where the assistant provides reasoning (either via a reasoning_content field or by virtue of the dataset’s design). This ensures every medical example teaches the model to “think before answering.”

🧾 Data Fields

• text (string): The full conversation formatted with the Llama 3.1 chat template, ready for language modelling.

🎨 Data Formatting

All conversations are formatted using the Llama 3.1 Instruct tokenizer’s apply_chat_template. Medical examples include a custom system prompt, and the assistant response is structured with <think> tags to separate reasoning from the final answer.

Example (Medical)

<|begin_of_text|><|start_header_id|>system<|end_header_id|>
Cutting Knowledge Date: December 2023
Today Date: 26 Jul 2024
You are Avicenna, an expert medical AI. Analyze the case step-by-step.<|eot_id|>
<|start_header_id|>user<|end_header_id|>
Are asymptomatic carriers a potential source for transmission of epidemic and nonepidemic Clostridium difficile strains among long-term care facility residents?

A.: yes
B.: no
C.: maybe<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>
<think>
... (detailed reasoning) ...
</think>

Yes, asymptomatic carriers can transmit both epidemic and non-epidemic strains.<|eot_id|>

Example (General)

<|begin_of_text|><|start_header_id|>system<|end_header_id|>
Cutting Knowledge Date: December 2023
Today Date: 26 Jul 2024
You are a helpful AI assistant.<|eot_id|>
<|start_header_id|>user<|end_header_id|>
What is the capital of France?<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>
The capital of France is Paris.<|eot_id|>

The default system preamble (Cutting Knowledge Date...) is automatically added by the tokenizer and is harmless – it keeps the format consistent with the model’s pre‑training.

🎯 Intended Use

• Fine‑tuning large language models (especially Llama‑3.1‑based architectures) for medical reasoning tasks.
• Teaching models to generate step‑by‑step reasoning before delivering a final answer (chain‑of‑thought).
• Improving performance on medical benchmarks such as MedQA, MMLU‑Medical, and MedMCQA.
• Research into clinical decision support and explainable AI in medicine.

🛠️ Creation Process

1. Loading: Each dataset was loaded using the Hugging Face datasets library.
2. Filtering: OpenMed datasets were filtered to keep only examples where at least one assistant message contains reasoning_content. Other datasets were kept as‑is (they are assumed to contain reasoning by design).
3. Sampling: From each source, a target number of examples were randomly sampled.
4. Formatting: Custom formatting functions wrapped each example in the Llama 3.1 chat template, inserting <think> tags around reasoning content.
5. Mixing: All formatted examples were concatenated and shuffled.
6. Splitting: A 95% train / 5% validation split was created.

❓ FAQ

Q: Why do examples contain Cutting Knowledge Date...?
A: This is automatically added by the Llama 3.1 tokenizer’s chat template. It does not affect training quality and keeps the input format aligned with the model’s pre‑training.

Q: Will the model learn to answer multiple‑choice questions only?
A: No – the dataset includes both medical (often multiple‑choice) and general (open‑ended) examples. The model learns to reason from the question content, not just the options. During inference, you can prompt naturally.

Q: How do I load the dataset?

from datasets import load_dataset
dataset = load_dataset("salihfurkaan/Medical-Reasoning-CoT-MegaMix")
train_data = dataset["train"]
val_data = dataset["validation"]

📜 License and Attribution

This dataset is a derivative of several publicly available sources. Each source has its own license; users must comply with all original licenses:

Source	License (as stated on HF)
OpenMed/*	Apache 2.0
Intelligent-Internet/II-Medical-Reasoning-SFT	Unknown
FreedomIntelligence/medical-o1-reasoning-SFT	Apache 2.0
teknium/OpenHermes-2.5	MIT

The curated dataset itself is released under the same terms as the original sources. By using this dataset, you agree to respect the licenses of all constituent datasets.

🙋 Contact

For questions or feedback, you can contact me.

Cite

If you use this dataset in your work, please cite the original sources and this derivative:

@misc{medical-reasoning-cot-megamix,
author = {Salih Furkaan},
title = {Medical-Reasoning-CoT-MegaMix},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/salihfurkaan/Medical-Reasoning-CoT-MegaMix}}
}