METRIC-BASED COMPARISON OF FINE-TUNED LLAMA 2 AND MIXTRAL LARGE LANGUAGE MODELS FOR INSTRUCTION TASKS
Bohdan M. Pavlyshenko, Ivan Bulka · Electronics and Information Technologies · 2024
The paper considers a comprehensive analysis and comparative study of two advanced Large Language Models (LLMs), namely LLaMA 2 and Mixtral, with a specific focus on their performance in executing instructional tasks. These models were fine-tuned using techniques such as LoRA and QLoRA, which were applied to extensive instruction datasets. The fine-tuning process was further enhanced by the implementation of Parameter-Efficient Fine-Tuning (PEFT) on NVIDIA A100 Tensor Core GPU instances, ensuring optimal performance. Both LLaMA 2 and Mixtral models were fine-tuned using the Hugging Face and PyTorch platforms, ensuring that similar parameters were maintained to facilitate a fair comparison. An inference was made using data not used in the initial training phase. This approach was adopted to test the models' ability to generalize and adapt to new, unseen data, thereby providing a more robust evaluation of their performance. An evaluation framework was established using the RAGAS library. The framework was designed to provide precise and reliable metrics, offering a comprehensive assessment of the models' performance. While the LLaMA 2 model demonstrates a faster rate of fine-tuning, it is susceptible to overfitting. On the other hand, Mixtrail, despite requiring more time for training, outperforms in evaluations, making it a more dependable tool for instructional tasks. Keywords: LLMs, PEFT, Lora, Qlora, Mixtral, LLaMA, LLMs fine-tuning