This project was developed for the SoICT Hackathon 2024 - Legal Document Retrieval competition. The challenge focused on retrieving relevant Vietnamese legal documents for a given set of queries, using advanced natural language processing (NLP) techniques.
The goal of the project was to fine-tune the BAAI/bge-m3 model to optimize retrieval accuracy for Vietnamese legal texts. The approach involved creating a triplet dataset for training, fine-tuning both the embedder and reranker components, and evaluating the model on the competition's Public Test Data.
- Fine-Tuned Model: Improved the BAAI/bge-m3 model's performance using a custom triplet dataset derived from labeled training data (query-document pairs).
- Advanced Retrieval Techniques: Enhanced both dense and re-ranking stages of the document retrieval pipeline to ensure high accuracy.
- Evaluation Success: Ranked Top 16 on the Public Test leaderboard based on the MRR@10 metric.