🧹 Medical-Reasoning-CoT-MegaMix-Clean

This is a cleaned version of the Medical-Reasoning-CoT-MegaMix dataset. The only modification is the removal of the static system metadata line that was automatically added by the Llama 3.1 tokenizer's apply_chat_template function.

🔄 What Was Removed?

From every example, we removed the exact prefix:

<|begin_of_text|><|start_header_id|>system<|end_header_id|> Cutting Knowledge Date: December 2023 Today Date: 26 Jul 2024

This metadata is static and identical across all examples. It provides no task-relevant information and was removed to create a cleaner, more minimal training format.

Before Cleaning:

<|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|>...

After Cleaning:

<|begin_of_text|><|start_header_id|>system<|end_header_id|> You are Avicenna, an expert medical AI. Analyze the case step-by-step.<|eot_id|>...

All other structure – including special tokens, system prompts, user queries, assistant responses, and <think> tags – remains identical to the original.

📊 Dataset Composition (Unchanged from Original)

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

🧠 Why Use This Cleaned Version?

Reason	Explanation
Matches inference format	When you deploy your model, you likely won't include the static date metadata. Training without it better aligns with inference.
Cleaner input	Removes irrelevant tokens that the model doesn't need to process.
Slightly more efficient	~100 fewer tokens per example saves minimal but non-zero compute.
No performance loss	The removed metadata is static and provides no task-relevant information.

🎯 When to Use Which Version

Use Case	Recommended Version
Fine-tuning for deployment (matching inference)	✅ Clean version (this one)
Maximum compatibility with Llama 3.1 defaults	Original version
Research on prompt formatting effects	Both (for comparison)
Teaching the model to ignore metadata	Clean version

📥 How to Load

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

🙋 Contact

For questions or feedback, you can contact me.

📜 License and Attribution

This dataset is a derivative of the same sources as the original. 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

If you use this dataset, please cite the original sources and this cleaned version:

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