CUAD Question Answering with LoRA Fine-Tuning

Fine-tuned DistilBERT with LoRA on CUAD, updating only 1–2% of parameters for contract clause QA.

In this project, DistilBERT is fine-tuned on the CUAD dataset (41 clause types) using LoRA (Low-Rank Adaptation) so that only 1–2% of parameters are trained.

Highlights:

  • Parameter-efficient fine-tuning for legal QA
  • Support for multiple clause types in contracts
  • Training, evaluation, and reproducibility scripts

This setup is ideal for experimenting with legal-domain QA while keeping compute requirements manageable.

View the repository