Skip to content

Latest commit

 

History

History
40 lines (25 loc) · 2.88 KB

contribution_guide.md

File metadata and controls

40 lines (25 loc) · 2.88 KB

Contribution Guide for StarkWizard

Prerequisites

Before contributing to StarkWizard, please note the following requirements:

  • GPU Requirements for Training Models: A GPU with at least 8 GB of VRAM is needed for training models.
  • CPU and Memory for Merging: Merging is performed on the CPU and requires approximately 8 GB of RAM.
  • Python and Conda Installation: It is highly recommended to install Python and Conda.
  • Virtual Environment: Creating a virtual environment (venv) is advisable due to the extensive and specific requirements of the project, including fixed versions of some libraries.

How to Contribute

If You Are Not Familiar with Machine Learning

You can still contribute by helping to expand the corpus of question and answer data used to train the model. The sources.txt document shows a variety of sources used, and some topics are not fully covered, leading to inaccuracies in the model generation.

  • Working on Topics: Go to the corpus_src/train directory to see the list of covered topics. Each CSV file represents a source and a topic. For example, alexandria_queue.csv contains training data on the topic of queues from the Alexandria project.
  • Contributing to a Topic: You can open any CSV file to see the existing data and decide if you can or want to contribute to the topic. You can use a new source or add to the existing sources.
  • Using a New Source: If you add data from a new source 'X', create a file named X_my_topic.csv and add 'X' to the sources.txt file.
  • Submitting Your Contribution: Push your file and create a Pull Request (PR). If it passes the tests, it will be integrated into the project and added to the model during the next training session.

If You Wish to Participate in Training

  • Uploading Models: Direct the uploads of the models you generate to your Hugging Face account to avoid overwriting the reference model. After testing, we may switch the model to our StarkWizard account.
  • Tasks Available: Optimize hyperparameters and parameters, such as learning rate, optimizer, or the Q and dropout of the implemented QLora.

If You Wish to Participate in Quantization

  • Supported Formats: Two formats are currently supported - AWQ and GGUF. GGUF is the only format supported on Apple Silicon and by LM Studio.
  • Notebooks: Two notebooks are available, each corresponding to one of the formats. Note that not all cells need to be run every time; some are for installing or compiling libraries.
  • Improvements: Generate variations of models with different parameters, such as multiple Q values.
  • Testing and Uploading: Similar to model training, test your models and upload them to your Hugging Face account. We will import them after testing.

Your contributions are vital to the growth and improvement of the StarkWizard project. We look forward to your valuable input and collaboration!