OPERA: A Framework for Efficient Data Pruning in Retrieval Models

In the realm of dense retrieval systems, fine-tuning on domain-specific data is crucial. However, not every training pair is equally beneficial for the learning process. OPERA emerges as a solution, focusing on data pruning to optimize both the efficiency and effectiveness of model adaptation. The framework begins with static pruning (SP), which filters out low-similarity query-document pairs. This method highlights a critical balance between quality and coverage: while ranking performance (NDCG) sees improvement, retrieval effectiveness (Recall) may suffer due to a lack of diverse queries. To address this issue, OPERA introduces a dynamic pruning (DP) approach that adjusts sampling probabilities for queries and documents during training. This ensures that high-quality examples are prioritized while still utilizing the complete training dataset. Evaluations across eight diverse datasets indicate that SP enhances ranking metrics compared to traditional fine-tuning (NDCG@10 +0.5%). In contrast, DP outperforms both ranking (NDCG@10 +1.9%) and retrieval (Recall@20 +0.7%) metrics, achieving an average rank of 1.38 across methods. Furthermore, these results are applicable to Qwen3-Embedding, a dense retriever based on LLM architecture, demonstrating the framework's versatility. Remarkably, DP achieves similar outcomes in under half the training time of standard methods.