This repository contains the code and data used to fine-tune the Llama2-Chat model using the 4-bit quantisation QLoRA (Quantization with Low Rank Approximation) PEFT technique on the OpenOrca dataset.
The OpenOrca-Clean dataset is a refined version derived from the original OpenOrca dataset.
The Llama2-OpenOrca-Clean dataset is tailored specifically for fine-tuning the Llama2-Chat model. It is derived from the OpenOrca-Clean dataset, further adapted to fit the llama prompt template. The dataset comprises a single column labeled "text," structured in the given format-
![Screenshot 2024-04-06 115410](https://private-user-images.githubusercontent.com/111623667/320187060-b9f35f1d-5bcf-4d26-9e34-efbd4df7f744.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3MTg4MTMxMTMsIm5iZiI6MTcxODgxMjgxMywicGF0aCI6Ii8xMTE2MjM2NjcvMzIwMTg3MDYwLWI5ZjM1ZjFkLTViY2YtNGQyNi05ZTM0LWVmYmQ0ZGY3Zjc0NC5wbmc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjQwNjE5JTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI0MDYxOVQxNjAwMTNaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT02NjcxNjYxZmUyNGE0ZDlhMDM0MmQ1NGY1OWMzYzBhNjU0YTM4OGM3Yzc3YjY0NTBiYTRlZTUyNTZkNWU4ZjgxJlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCZhY3Rvcl9pZD0wJmtleV9pZD0wJnJlcG9faWQ9MCJ9.e38HF--b5K6hgVcKWsYgqrc8afhSWpymeu_OX6tE2vA)
- Base Model: Llama-2-7B-Chat-hf
- Fine-tuning Technique: 4-bit quantization using QLoRA PEFT
- Dataset Used: Llama2-OpenOrca-Clean
The fine-tuning process involves training the Llama2-Chat model with 4-bit quantization using the QLoRA technique. This technique allows for efficient representation of model parameters while minimizing computational overhead.
Llama-2-7B-Chat-OpenOrca
Our latest model, fine-tuned with 1000 examples using 4-bit quantization QLoRA from Llama2-OpenOrca-Clean dataset, is now available.