Unsloth LoRA Fine-tuning System

A modular LoRA fine-tuning pipeline built on Unsloth for sentiment analysis, with 4-bit quantization, early stopping, and adaptive Llama-2/Llama-3 support.

Overview

A comprehensive LoRA (Low-Rank Adaptation) fine-tuning system built on Unsloth for sentiment-analysis tasks. It covers the full lifecycle — training, evaluation, and interactive testing — and adapts automatically to different base models (Llama-2 / Llama-3).

Implementation

Efficient training: Unsloth optimization for ~2x faster training, with 4-bit quantization so large models fit on a single consumer GPU (~7GB VRAM)
LoRA config: rank-32 adapters on attention layers, cosine LR scheduling, ~100MB adapter output
Early stopping: custom callback with patience/threshold to prevent overfitting and cut training time
Modular architecture: clean separation across config / data_processor / model_manager / training_pipeline / evaluator / model_saver, orchestrated by a single lora_system controller
Dataset: IMDb movie reviews (25k samples, 80/20 split), extensible to other Hugging Face datasets
Tooling: interactive & automated training launchers, model version management, and a free-form text tester returning predictions with confidence scores
Result: ~85–90% accuracy on the IMDb test set, ~15–30 min training on an RTX 4090