Multi-Strategy Fine-Tuning of LLMs for Sequence Classification: BERT and LLaMA with Soft and Hard Prompting

Standard Fine-tuning with BERT:

• Model: bert-base-uncased
• Dataset: GLUE MRPC dataset
• Layers: All BERT layers are initially frozen, except for the last two encoder layers, which are unfrozen and fine-tuned.
• Optimizations: Learning rate, batch size, number of epochs, weight decay, and other hyperparameters were tuned for optimal performance.
• Evaluation Metrics: Accuracy, classification report, confusion matrix, and ROC-AUC curve.
• Output: The fine-tuned model and tokenizer are saved for further inference.
•  

Soft Prompting:

• Instead of adding tokens explicitly as in hard prompting, this method introduces trainable soft prompts (embedding vectors) which are prepended to the input token embeddings.
• The soft prompt model concatenates a learnable soft prompt with the original input embeddings and uses BERT for sequence classification.
• Evaluation: Accuracy, classification report, confusion matrix, and ROC-AUC were used to evaluate model performance.
• Output: The soft-prompting model and tokenizer are saved.
•  

Hard Prompting with LLaMA:

• Hard prompts explicitly structure the task within the input sentences. For instance, the task “Does sentence 1 mean the same as sentence 2?” is embedded into the input directly.
• Model: LLaMA 3.1 was used for inference with temperature control for deterministic output.
• Performance was evaluated using classification accuracy and reports.
•  

Evaluation and Analysis:

• The final evaluation was performed using multiple metrics, such as classification reports, confusion matrices, and ROC-AUC curves for each model.
• Visualization tools (matplotlib) were used to plot confusion matrices and ROC curves, allowing easy comparison of model performance.