Fine-Tuning Llama 3 8B for High-Accuracy Sentiment Analysis

This project documents the process of fine-tuning the meta-llama/Meta-Llama-3-8B-Instruct model to transform it from a general-purpose instruction model into a specialized sentiment analysis classifier.The baseline model exhibited significant weaknesses, including a strong positive bias and poor instruction adherence. Through targeted fine-tuning using QLoRA, this project successfully corrected these flaws, resulting in a robust and accurate model. Fine-Tuned Model on Hugging Face HubThe final LoRA adapters for this project are publicly available on the Hugging Face Hub. You can access and use it here

Performance: Before vs. After

The primary goal of this project was to demonstrably improve upon the zero-shot performance of the base model. The results show a dramatic increase in both accuracy and the model's ability to correctly classify negative sentiment.

• Baseline (Zero-Shot):
  • Overall Accuracy: 50.95%
  • F1-Score (Negative): 0.01
• Fine-Tuned (QLoRA):
  • Overall Accuracy: 82.30%
  • F1-Score (Negative): 0.78

Key Components:

• Model: meta-llama/Meta-Llama-3-8B-Instruct
• Dataset: A 5,000-sample subset of the Amazon Polarity dataset.
• Fine-Tuning Technique: QLoRA (Quantized Low-Rank Adaptation) to train the model in 4-bit precision.
• Optimization Library: Unsloth, which provided a speedup and significant memory savings over standard Hugging Face libraries.
• Evaluation: An evaluation pipeline was built to parse the models' outputs and compare the baseline performance against the final fine-tuned version on a 1,000-sample hold-out test set.