Llama 2 Fine-Tuning on Databricks Dolly 15k Subset

The code in this repository fine-tunes a Llama 2 model on a 1000-sample subset of the Databricks Dolly 15k instruction dataset using Supervised Fine-Tuning (SFT) with QLoRA 4-bit precision.

1. Clone this repository:

   git clone https://github.com/golkir/llama2-7b-minidatabricks.git
   cd llama2-7b-minidatabricks
   

2. Install dependencies

   pip install .

3. Run the dataset subset creation script which fetches the Dolly 15k dataset and processes it in Llama 2 instruction format.

   python load-databricks.py
   

4. Run the fine-tuning script:

   python finetuning.py

• The Dolly 15k dataset is originally provided by Databricks. Link to Databricks Dolly 15k dataset.
• The Llama 2 model can be found in HuggingFace repository.

This code is licensed under the Apache 2.0 License.