This repository contains a version of Phi-2 fine tuned specifically for coding tasks.
This model builds upon Phi-2 by Microsoft and it is fine tuned on CodeAlpaca-20k dataset. This model is fine tuned for coding purposes.
Phi-2 is not a RLHF tuned LLM. This fine tuning also adds small chat capabilities to the LLM using Human: <prompt> Assistant:
paradigm.
- Enhanced Domain Specificity: Improved accuracy and relevance in coding.
- Implementation of QLoRA: Used QLoRA with 4-bit quantization for fine tuning
To use this fine-tuned model, ensure you have the following:
- Python 3.8 or later
- Run the cells in the notebook
- Clone the repository:
git clone https://github.com/Nabeegh-Ahmed/phi-2-coding-expert.git
- Run the last two cells of the notebook for inference
- I had very limited access to GPUs so fine tuning even on a small dataset took a huge amount of time.
- For the same reason, I was not able to fine tune on the whole dataset with short batch sizes. I only fine tuned on a small dataset to prevent CUDA out of memory errors.
This project is licensed under the terms of the MIT license.
This model is based on the work of Microsoft. We thank them for their foundational contributions to the field of AI and language understanding.