How I Reproduced BM25, Dense Retrieval, and SPLADE on a 16GB MacBook

A hands-on reproduction of three information retrieval baselines — BM25, dense retrieval (BGE-base-en-v1.5), and SPLADE v3 — on a 16GB MacBook using the Castorini onboarding path with Anserini and Pyserini. The post covers setup friction not documented officially: shallow git clones to avoid timeouts, Java 21 version pinning, a memory crash requiring heap tuning for MS MARCO indexing, and a segfault from OpenMP library conflicts in dense retrieval. All expected scores were reproduced exactly. A dot-product verification script shows how BM25 scores can be manually audited. On NFCorpus, BM25 scored 0.3218 nDCG@10, dense retrieval 0.3808, and SPLADE 0.3624, with SPLADE requiring ~68 minutes of CPU encoding vs ~4 minutes for BGE-base. Practical advice for RAG builders: verify retrieval baselines before tuning prompts or swapping models.

Nguồn: https://towardsdatascience.com/how-i-reproduced-bm25-dense-retrieval-and-splade-on-a-16gb-macbook. 8sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.