lora-tuning-text

LoRA Fine-Tuning — GPT-2 on SQuAD Q&A

Fine-tunes gpt2 on the SQuAD dataset using LoRA (Low-Rank Adaptation) via the PEFT library. Trained on MacBook Pro M2 using the MPS backend.

Problem

Stock GPT-2 was trained for text continuation, not Q&A. When asked a direct question it wanders and generates off-topic text. LoRA fine-tuning on SQuAD teaches it the Question → Answer response shape.

Dataset

Method — LoRA

LoRA injects small trainable rank-decomposition matrices into the attention layers, keeping the base model frozen.

• ~0.1% of total parameters are trained
• Only the adapter weights are saved (~few MB vs ~500 MB for full GPT-2)
• GPT-2 uses a combined QKV projection (c_attn), so LoRA targets that single module

Project Structure

lora-tuning-text/
├── lora_train.py        # Training script
├── test_base_model.py   # Inference with base GPT-2 (no fine-tuning)
├── test_tuned_model.py  # Inference with the LoRA-tuned adapter
├── requirements.txt
└── lora-gpt2-squad/     # Output directory (auto-generated)
    └── final/           # Saved LoRA adapter

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Usage

Step 1 — See how bad base GPT-2 is:

python test_base_model.py

Step 2 — Fine-tune with LoRA:

python lora_train.py

Step 3 — Compare with the tuned model:

python test_tuned_model.py

Output

• Local adapter: ./lora-gpt2-squad/final/

Notes

• Both test scripts use the same 10 SQuAD-style factual questions for a fair comparison
• Training takes ~20–30 minutes on M2 MPS for 3 epochs over 8k examples

---

Result

We partially achieved the result. The LoRA-tuned GPT-2 did learn the Question → Answer response shape and stopped wandering as much as the base model. However, it still occasionally hallucinated or gave incomplete answers — full factual accuracy was not consistently reached.