LoRA Hypernetwork for Fact Learning

This project trains a hypernetwork to generate LoRA adapters for teaching new facts to language models.

What it does

Instead of manually training a LoRA for each new fact, the hypernetwork learns to predict what LoRA weights are needed. You give it a fact, and it outputs the adapter weights.

Files

• train_lora_adapters.py - Creates individual LoRA adapters for each fact in the dataset
• train_hypernetwork.py - Trains the hypernetwork to predict LoRA weights
• test_hypernetwork.py - Tests the trained hypernetwork
• evaluate_real_loras.py - Evaluates the individually trained LoRAs
• check_base_knowledge.py - Checks if the base model already knows certain facts

Data

Uses wikifactdiff_converted.csv which contains questions and updated answers for fact changes.

Usage

First, generate the training data (individual LoRAs):

python train_lora_adapters.py --num_facts 100

Then train the hypernetwork:

python train_hypernetwork.py

Test it:

python test_hypernetwork.py

Model

Base model: Qwen2-7B-Instruct LoRA config: rank=2, alpha=4, targets=q_proj,v_proj