NITA_R_DPR — Retrieval Only Task

Is this a manual (human intervention) or automatic run?: automatic
Does this run leverage neural networks?: yes
Does this run leverage proprietary models in any step of the retrieval pipeline?: no
Does this run leverage open-weight LLMs (> 5B parameters) in any step of the retrieval pipeline?: no
Does this run leverage smaller open-weight language models in any step of the retrieval pipeline?: yes
Was this run padded with results from a baseline run?: no
What would you categorize this run as?: Multi-Stage Pipeline pointwise
Please provide a short description of this run: This submission implements a three-stage hybrid retrieval pipeline for TREC RAG 2025. Stage 1 utilizes pre-computed BM25 results to select the top 500 lexical candidates per query. Stage 2 applies DPR semantic filtering using Facebook's dpr-question_encoder-single-nq-base and dpr-ctx_encoder-single-nq-base models to reduce candidates to the top 200 based on cosine similarity of dense embeddings. Stage 3 performs neural reranking with cross-encoder/ms-marco-MiniLM-L-12-v2 for final relevance scoring. The system processes 105 queries with GPU batch inference (batch size 256), delivering top-100 results that combine lexical matching with semantic understanding through transformer-based architectures.
Please give this run a priority for inclusion in manual assessments.: 3