Update README.md

# DeepSeek-R1-Distill-Llama-8B-Medical-COT

## 🏥 Fine-tuned Medical Model
This is a **fine-tuned version of DeepSeek-R1-Distill-Llama-8B**, optimized for **medical reasoning and clinical case analysis** using **LoRA (Low-Rank Adaptation) with Unsloth**.

- **Base Model:** [DeepSeek-R1-Distill-Llama-8B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)
- **Fine-Tuning Framework:** [Unsloth](https://github.com/unslothai/unsloth)
- **Dataset:** [FreedomIntelligence/medical-o1-reasoning-SFT](https://huggingface.co/datasets/FreedomIntelligence/medical-o1-reasoning-SFT)
- **Quantization:** 4-bit (bitsandbytes)
- **Task:** **Clinical reasoning, medical question-answering, diagnosis assistance**

---

## 📖 Model Details

| Feature | Value |
|--------------------|-------------|
| **Architecture** | Llama-8B (Distilled) |
| **Language** | English |
| **Training Steps** | 60 |
| **Batch Size** | 2 (with gradient accumulation) |
| **Gradient Accumulation Steps** | 4 |
| **Precision** | Mixed (FP16/BF16 based on GPU support) |
| **Optimizer** | AdamW 8-bit |
| **Fine-Tuned With** | PEFT + LoRA (Unsloth) |

---

## 📊 Training Summary
**Loss Trend During Fine-Tuning:**

| Step | Training Loss |
|------|--------------|
| 10 | 1.9188 |
| 20 | 1.4615 |
| 30 | 1.4023 |
| 40 | 1.3088 |
| 50 | 1.3443 |
| 60 | 1.3140 |

---

## 🚀 How to Use

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "develops20/DeepSeek-R1-Distill-Llama-8B-Medical-COT"

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Run inference
def ask_model(question):
inputs = tokenizer(question, return_tensors="pt").to("cuda")
outputs = model.generate(input_ids=inputs.input_ids, max_new_tokens=512)
return tokenizer.decode(outputs[0], skip_special_tokens=True)

question = "A 61-year-old woman has involuntary urine loss when coughing. What would cystometry likely reveal?"
print(ask_model(question))
```

---

## 📌 Example Outputs

### **Q:** "A 59-year-old man presents with fever, night sweats, and a 12mm aortic valve vegetation. What is the most likely predisposing factor?"
**🔹 Model's Answer:** "The most likely predisposing factor for this patient’s infective endocarditis is a history of **valvular heart disease** or **prosthetic valves**, given the presence of an aortic valve vegetation. The causative organism is likely **Enterococcus species**, which does not grow in high salt concentrations."

---

## 🏆 Why Use This Model?
✅ **Fine-tuned on a structured medical reasoning dataset** 🔬
✅ **Optimized for speed with Unsloth** ⚡
✅ **Lower VRAM usage via 4-bit quantization** 🏗️
✅ **Handles medical Q&A, diagnosis reasoning, and case analysis** 🏥

---

## 🔧 Fine-Tuning Details
This model was fine-tuned using **Parameter Efficient Fine-Tuning (PEFT) with LoRA** in **Unsloth**, allowing efficient adaptation without full model training.

### **Training Arguments:**
```python
TrainingArguments(
per_device_train_batch_size=2,
gradient_accumulation_steps=4,
num_train_epochs=1,
learning_rate=2e-4,
warmup_steps=5,
max_steps=60,
optim="adamw_8bit",
weight_decay=0.01,
lr_scheduler_type="linear",
fp16=True, # BF16 if supported
output_dir="outputs"
)
```

---

## 📜 License & Contribution
**License:** MIT

✅ Feel free to use, modify, and improve this model. If you use it in research or projects, consider citing this work!

**Contribute & Feedback:** If you have suggestions or improvements, please open an issue or pull request on Hugging Face.

---

## 🤝 Acknowledgments
This model was trained with the support of **Kaggle's free GPUs** and the **Hugging Face Transformers ecosystem**. Special thanks to the **Unsloth** developers for optimizing LoRA fine-tuning!

README.md

@@ -1,7 +1,13 @@
----
-license: apache-2.0
-tags:
-- unsloth
-- trl
-- sft
----
+---
+license: apache-2.0
+tags:
+- unsloth
+- trl
+- sft
+datasets:
+- FreedomIntelligence/medical-o1-reasoning-SFT
+language:
+- en
+base_model:
+- deepseek-ai/DeepSeek-R1-Distill-Llama-8B
+---