CCA GPT Model Trainer

This repository contains a script created specifically for my employer, designed with MediaWiki and Jira Cloud in mind. The tool automates the process of fetching, cleaning, and combining data from MediaWiki and Jira, and then fine-tuning a language model (GPT) on the combined dataset. Currently, all settings are hardcoded in the Python code, but I will be changing that soon. The script is optimized to handle GPU memory constraints and can switch to CPU if needed.

Requirements

• Python 3.x
• Required Python packages (install using pip install -r requirements.txt):
  • art
  • colorama
  • MySQLdb
  • beautifulsoup4
  • requests
  • transformers
  • torch
  • datasets

Usage

1. Clone the repository:

   git clone https://github.com/yourusername/cca-gpt-model-trainer.git
   cd cca-gpt-model-trainer

2. Install dependencies:

   pip install -r requirements.txt

3. Update settings in the script:

   Open the cca-gpt-model-trainer.py script and update the following settings:

   • MySQL Database Connection:

     connection = MySQLdb.connect(
         host="localhost", user="grahf", password="<password>", database="local_wiki"
     )

     Change host, user, password, and database to match your MediaWiki database credentials.

   • Jira API Connection:

     url = "https://site.atlassian.net/rest/api/3/search"  # CHANGE THIS TO APPROPRIATE JIRA URL
     params = {
         "jql": "project = CSS",  # CHANGE THIS TO APPROPRIATE PROJECT CODE
         "maxResults": 3000,
         "fields": "summary,description,comment",
     }
     
     email = "[email protected]"  # CHANGE THIS TO APPROPRIATE JIRA USER
     api_token = "<token>"  # ADD JIRA TOKEN

     Change the url, params['jql'], email, and api_token to match your Jira Cloud instance and credentials.

4. Run the script:

   python cca-gpt-model-trainer.py

Script Functions

• fetch_mediawiki_data(): Fetches MediaWiki data and saves it to a text file. Ensure you update the database connection settings.
• clean_mediawiki_data(): Cleans the MediaWiki data by removing HTML tags and other unnecessary content.
• fetch_jira_data(): Fetches Jira entries using the Jira REST API and saves them to a JSON file. Ensure you update the Jira connection settings.
• combine_files(jira_file, mediawiki_file, combined_file): Combines Jira and MediaWiki data into a single text file for training.
• download_tokenizer_files(model_name, output_dir): Downloads the tokenizer files for the specified model.
• fine_tune_gpt_model(data_file, output_dir): Fine-tunes a GPT model on the combined data.

Future Improvements

• Externalize settings to a configuration file to avoid hardcoding values in the script.
• Add logging for better traceability and debugging.
• Improve error handling and retry mechanisms.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Author

• Dean Thomson (grahfmusic) - GitHub