Gregniuki/polish-medical-cot-PES

Polish Medical Chain-of-Thought Dataset (PES / LEK / LDEK)

📌 Dataset Overview

polish-medical-cot-PES is a high-quality dataset of 33,774 Polish medical question-answering examples paired with detailed Chain-of-Thought (<think> ... </think>) clinical reasoning.

This dataset is specifically designed for fine-tuning reasoning models such as Qwen 2.5, DeepSeek-R1-Distill-Qwen, Llama 3, and Mistral on Polish medical knowledge, medical licensing exams (LEK / LDEK / PES), and step-by-step clinical decision-making.

---

💡 Data Format

The dataset is formatted in standard OpenAI / Hugging Face Chat JSONL structure (messages array).

{
  "messages": [
    {
      "role": "system",
      "content": "Jesteś ekspertem medycznym i tutorem specjalizującym się w polskich egzaminach lekarskich..."
    },
    {
      "role": "user",
      "content": "### Question ID: 81
**Specialization:** Alergologia
**Question:** Reakcja nadwrażliwości typu II nie jest mechanizmem..."
    },
    {
      "role": "assistant", "content": "<think>
Typy nadwrażliwości według klasyfikacji Gella-Coombsa. Typ II to nadwrażliwość cytotoksyczna...
</think>

Analiza poszczególnych opcji...

Prawidłowa odpowiedź to: B"
    }
  ]
}

---

🧹 Quality Assurance & Cleaning Pipeline

The dataset underwent rigorous automated verification:

• Total Initial Records: 35,005
• Valid Clean Records: 33,774 (96.48%)
• Removed Bad Records: 1,231 (3.52%)

Rejection Breakdown:

• Missing / Unclosed <think> Tags: 973 records removed.
• Short / Empty Responses: 153 records removed.
• Truncated / Cut-off Endings: 71 records removed.
• Model Degeneration Loops: 31 records removed.
• Missing Assistant Content / Corrupted Lines: 3 records removed.

---

🚀 How to Use & Fine-Tune

1. Load with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("Gregniuki/polish-medical-cot-PES", split="train")
print(dataset[0])

2. Fine-Tuning Qwen with TRL / Unsloth

from datasets import load_dataset
from trl import SFTTrainer
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Qwen/Qwen2.5-7B-Instruct",
    max_seq_length=2048,
    load_in_4bit=True,
)

dataset = load_dataset("Gregniuki/polish-medical-cot-PES", split="train")

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="messages",
    max_seq_length=2048,
)

trainer.train()

---

📜 License

This dataset is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.