README.md

---
{}
---

# CodeGPT Fine-tuned for Code Generation

## Model Description
This model is a fine-tuned version of [microsoft/CodeGPT-small-py](https://huggingface.co/microsoft/CodeGPT-small-py) 
trained on coding problems and solutions for code generation tasks.

## Training Details
- **Base Model:** microsoft/CodeGPT-small-py (124M parameters)
- **Dataset:** Rabinovich/Code-Generation-LLM-LoRA (500 examples)
- **Epochs:** 2
- **Learning Rate:** 5e-5
- **Batch Size:** 4
- **Hardware:** CPU

## Training Results
| Step | Training Loss |
|------|--------------|
| 25   | 4.4322       |
| 50   | 3.4648       |
| 100  | 3.1430       |
| 150  | 2.7050       |
| 200  | 2.7491       |
| 250  | 2.7126       |

**Loss improved from 4.43 → 2.71 (39% reduction)**

## How to Use
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Pradnya27/codegpt-finetuned-code-generation")
tokenizer = AutoTokenizer.from_pretrained("Pradnya27/codegpt-finetuned-code-generation")

prompt = "Generate code: Write a function to check if a number is prime"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(inputs["input_ids"], max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Limitations
- Trained on a small subset (500 examples) — larger training would improve results
- Works best with competitive programming style problems
- Output quality improves with more specific prompts

## Future Work
- Train on full dataset (34,727 examples)
- Experiment with LoRA fine-tuning
- Evaluate on HumanEval benchmark