Fine-tuning Phi-4-Mini-Instruct with LoRA on Resume Question Answers.

Overview

This script fine-tunes the microsoft/Phi-4-mini-instruct model using LoRA (Low-Rank Adaptation) on a custom QA dataset (cleaned_qa.json). The model is quantized to 8-bit for efficient training.

Setup

1. Install dependencies:

   pip install -r requirements.txt

2. Place the QA dataset (cleaned_qa.json) in the working directory.

Training Steps

• Loads the Phi-4-mini-instruct model with 8-bit quantization.
• Formats the dataset to match the model's prompt style.
• Applies LoRA for parameter-efficient fine-tuning.
• Trains the model for 30 epochs with a batch size of 2.
• Saves the fine-tuned model and tokenizer to ./phi4_mini_finetuned.

Running the Script

Execute the script to start training:

python train.py

Output

• Fine-tuned model and tokenizer saved in ./phi4_mini_finetuned.
• Training logs and checkpoints stored in the output directory.

Notes

• Adjust hyperparameters (learning_rate, num_train_epochs, etc.) in TrainingArguments as needed.
• Ensure the dataset follows the expected JSON format with question and answer fields.