DeepSeek‑R1‑Medical‑COT

DeepSeek‑R1‑Medical‑COT is a 4‑bit fine‑tuned causal language model optimised for medical reasoning and clinical‑scenario interpretation.
It is based on unsloth/DeepSeek‑R1‑Distill‑Llama‑8B and fine‑tuned on the FreedomIntelligence/medical-o1-reasoning-SFT dataset to provide structured, step‑by‑step clinical reasoning and evidence‑based conclusions.

---

Repository contents

├── README.md

├── Fine_Tune_DeepSeek_R1_Medical_COT.ipynb # notebook with full training workflow

└── requirements.txt # Python dependencies

The Jupyter notebook contains all of the steps: model loading, prompt formatting, data preparation, LoRA configuration, training, validation, saving and pushing to the Hugging Face hub.

---

Requirements

Create a virtual environment and install the required packages:

pip install unsloth trl peft accelerate bitsandbytes datasets
pip install transformers gradio huggingface-hub

---

Usage

Load and run the model as shown below:

from unsloth import FastLanguageModel
from transformers import AutoTokenizer

model_name = "DeepSeek-R1-Medical-COT"

# load model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name, load_in_4bit=True
)

prompt = """
### Clinical Scenario:
A 54-year-old man complains of frequent urinary urgency, nocturia,
and a weak urinary stream. His prostate is moderately enlarged.
Predict likely cystometric findings.
"""

inputs = tokenizer([prompt], return_tensors="pt").to("cuda")
outputs = model.generate(
    input_ids=inputs.input_ids,
    attention_mask=inputs.attention_mask,
    max_new_tokens=500
)
print(tokenizer.decode(outputs[0]))

---

Training

Follow the instructions in Fine_Tune_DeepSeek_R1_Medical_COT.ipynb:

• Load unsloth/DeepSeek-R1-Distill-Llama-8B with load_in_4bit=True.
• Define a CoT‑style prompt template for medical questions.
• Load the FreedomIntelligence/medical-o1-reasoning-SFT dataset and format examples.
• Apply LoRA to selected attention and feed‑forward modules.
• Configure TrainingArguments (batch 1 + accumulation, 4‑bit AdamW, FP16/BF16, 200 steps, etc.).
• Train with trl.SFTTrainer.
• Validate on a held‑out slice, save validation_results.csv and training_logs.csv.
• Save the fine‑tuned model locally and push to the Hugging Face hub.
• You can run the notebook interactively with Jupyter or convert it to a script with jupyter nbconvert.

---

Validation

The notebook creates a small validation split (train[500:550]), runs inference on a few examples, and stores outputs/lengths.

---

Pushing to Hugging Face

Set your token in an environment variable:

export HF_TOKEN="your_token_here"

Then run (see notebook for full code):

from huggingface_hub import login, HfApi
import os

login(token=os.environ["HF_TOKEN"])
api = HfApi()
api.create_repo("DeepSeek-R1-Medical-COT", repo_type="model",
                exist_ok=True)
model.push_to_hub("DeepSeek-R1-Medical-COT",
                  tokenizer=tokenizer,
                  save_method="merged_16bit")

After this the model can be loaded with FastLanguageModel.from_pretrained("DeepSeek-R1-Medical-COT").

---

Simple Gradio demo

The notebook includes an example Gradio interface that accepts a medical scenario and returns step‑by‑step reasoning plus a final conclusion.

---

Model details

• Model type: Causal Language Model (LLM)
• Base model: unsloth/DeepSeek-R1-Distill-Llama-8B
• Fine‑tuned for: Medical instruction‑following and clinical reasoning

---

Sources

• Model hub: https://huggingface.co/MohamedASAK/DeepSeek-R1-Medical-COT
• Training dataset: https://huggingface.co/datasets/FreedomIntelligence/medical-o1-reasoning-SFT

---

Training details

• Dataset: FreedomIntelligence/medical-o1-reasoning-SFT
• Preprocessing: Prompts formatted with CoT style (grounded) for step-by-step reasoning
• Fine‑tuning method: LoRA applied to attention and feedforward modules
• Hyperparameters:
  • Batch size: 1 (gradient accumulation 8)
  • Max steps: 200
  • Learning rate: 2 × 10⁻⁴
  • Mixed precision: FP16 / BF16 depending on GPU support
  • Optimizer: 8‑bit AdamW

---

Evaluation

• Evaluated on a subset of medical reasoning questions
• Metrics: correctness of reasoning, coherence, and answer accuracy
• Results indicate improved structured reasoning over the base model

---

Limitations and Risks

• Limited to the quality and scope of the training dataset
• May not cover rare or highly specialized medical cases
• Should not replace clinical judgment; intended for educational and reasoning support

Recommendation: Always review model outputs with a qualified healthcare professional.