/fine-tune-with-unsloth

Primary LanguagePython

Fine-Tuning with Unsloth

This project demonstrates fine-tuning and inference using the Qwen2.5-Coder model with the unsloth library.

Setup

  1. Install required libraries: 
    pip install unsloth transformers datasets trl peft
  2. Authenticate Hugging Face: Set your Hugging Face token: 
    from google.colab import userdata
HF_TOKEN = userdata.get('HF_TOKEN')

Steps

1. Initialize the Model

The script loads the model and tokenizer with LoRA adapters for efficient fine-tuning.

2. Prepare the Dataset

It uses an Alpaca-style dataset for training (iamtarun/python_code_instructions_18k_alpaca).

3. Train the Model

The fine-tuning process is configured to run for 2 epochs with a batch size of 2.

4. Save and Share

After training:

  • Optionally push it to the Hugging Face Hub.

5. Run Inference

Use the model to generate responses for custom instructions. Example:

results = Inference.run_inference(
    model, tokenizer, 
    "Write a function for a merge interval.", ""
)
print(results[0])

Notes

  • Ensure CUDA is enabled for better performance.
  • Update ORG_NAME and MODEL to use your Hugging Face repository.

That's it! You can now fine-tune.