https://github.com/alexisbalayre/uk-resume-llm
Resume Insight: Fine-Tuning LLM for Resume Analysis
https://github.com/alexisbalayre/uk-resume-llm
fine-tuning llm machine-learning nlp
Last synced: 7 months ago
JSON representation
Resume Insight: Fine-Tuning LLM for Resume Analysis
- Host: GitHub
- URL: https://github.com/alexisbalayre/uk-resume-llm
- Owner: AlexisBalayre
- Created: 2024-02-27T08:31:25.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2024-03-01T01:12:44.000Z (over 1 year ago)
- Last Synced: 2025-03-29T11:12:33.272Z (7 months ago)
- Topics: fine-tuning, llm, machine-learning, nlp
- Language: Python
- Homepage:
- Size: 227 KB
- Stars: 20
- Watchers: 1
- Forks: 5
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Resume Insight: Fine-Tuning LLM for Resume Analysis
Resume Insight is a sophisticated project designed to harness the power of large language models (LLMs) for deep analysis and understanding of resume data. It seamlessly blends technologies like OCR, natural language processing (NLP), and machine learning (ML) to extract, structure, and generate insightful question-answer pairs from resumes. This data then serves as a foundation for fine-tuning an LLM, aiming to produce a model adept at generating nuanced, context-aware responses based on resume content.
## Key Features
- **Comprehensive Data Extraction**: Utilises OCR to transform PDF resumes into analysable text.
- **Intelligent Data Structuring**: Applies NLP techniques to categorise and structure resume text into meaningful data points.
- **Dynamic Dataset Creation**: Generates a rich dataset of question-answer pairs tailored for LLM training, focusing on resume insights.
- **Advanced Model Fine-Tuning**: Employs cutting-edge fine-tuning strategies, including memory optimisation and parameter-efficient fine-tuning (PEFT), to enhance an LLM's capability in generating contextually relevant text.
## Getting Started
### Prerequisites
- Python 3.8+
- PyTorch compatible with your CUDA version
- Hugging Face Transformers, Datasets
- pdf2image, pytesseract for OCR
- Spacy for NLP tasks
- bitsandbytes for memory optimisation
- Ollama to generate the training dataset
### Installation
First, clone this repository to your local machine:
```bash
git clone https://AlexisBalayre/UK-Resume-LLM.git
cd UK-Resume-LLM
```
Install the required Python packages:
```bash
pip install -r requirements.txt
```
Ensure you have Tesseract-OCR and the necessary language data installed for `pytesseract`.
### Preparing Your Dataset
1. **Generate Q&A Pairs**: Create a dataset of question-answer pairs based on the structured data. You have to install [ollama](https://ollama.com/) to generate the training dataset. Don't forget to modify the path of your UK Resume in `generate_training_dataset.py`.
```bash
ollama run mistral
python generate_training_dataset.py
```
2. **Prepare Training and Validation Sets**: Shuffle and split the generated dataset into training and validation sets.
```bash
python split_training_dataset.py
```
### Fine-Tuning the Language Model
To start the fine-tuning process:
```bash
python train_llm.py
```
This script will train the model on your custom dataset, leveraging the previously generated question-answer pairs.
## Converting and Quantising the Model
1. Clone the [llama.cpp](https://github.com/ggerganov/llama.cpp) repository and build the `llama` binary.
2. Convert the fine-tuned model to the `gguf` format:
```bash
cd /path/to/llama.cpp
python3 convert.py --vocab-type hfft /path/to/your/fine-tuned-model-directory
```
3. Quantise the model to a lower precision:
```bash
./quantize /path/to/your/fine-tuned-model-directory/ggml-model.gguf /path/to/your/fine-tuned-model-directory/ggml-model-Q4_K_M.gguf Q4_K_M
```
4. Finally, you can test the quantised model using [ollama](https://ollama.com/), [LM Studio](https://lmstudio.ai/) or any other compatible tool. For exemple with `ollama`, modify the path to the quantised model in the `Modelfile` file and run the following command:
```bash
ollama create your-model-name -f Modelfile
ollama run your-model-name
```
## Customization
Feel free to adjust the scripts to better fit your needs. For example, you might want to:
- Modify regular expressions in `uk_resume_data_extraction.py` for improved data extraction.
- Tweak the training parameters in `train_llm.py` to optimise model performance.
## Contributing
We welcome contributions! If you have suggestions for improvements or new features, please feel free to fork the repository, make your changes, and submit a pull request.
## License
This project is open-source and available under the [MIT License](LICENSE.md).
## Acknowledgments
This project makes extensive use of the following libraries and frameworks:
- [Hugging Face Transformers](https://github.com/huggingface/transformers) for NLP model support.
- [pdf2image](https://github.com/Belval/pdf2image) and [pytesseract](https://github.com/madmaze/pytesseract) for PDF text extraction.
- [Spacy](https://spacy.io/) for advanced NLP processing.
- [bitsandbytes](https://github.com/TimDettmers/bitsandbytes) for model memory optimization.
- [Mistral Mastery: Fine-Tuning & Fast Inference Guide](https://medium.com/@parikshitsaikia1619/mistral-mastery-fine-tuning-fast-inference-guide-62e163198b06)
- [Mistral-7B Fine-Tuning: A Step-by-Step Guide](https://gathnex.medium.com/mistral-7b-fine-tuning-a-step-by-step-guide-52122cdbeca8)
Your feedback and contributions are highly appreciated as we aim to continuously improve and expand the capabilities of Resume Insight.