baselines/msmarco_scale/README.md

Baseline notes — full MS MARCO / TREC-DL19

run_bm25_test_streaming_effectiveness.py is a local effectiveness sanity check over all 8,841,823 passages. It is deliberately not reported as an indexed BM25 latency measurement because it scans the corpus.

BASELINE_COMPARISON.json records the locally measured frozen-system result and published same-test reference rows. Published rows should be cited to NIST/Pyserini/the original dense-retrieval papers in manuscripts.

TREC-DL19 passage judgments require two evaluation conventions:

• nDCG@10: retain graded qrel values (trec_eval ndcg_cut uses raw relevance values as gains);
• binary MRR/P/Recall/Hit: grade 1 is only Related, not relevant, therefore use relevance >= 2.

The local frozen-system evaluator in ../scale_experiments/msmarco_test_baseline_compare.py returns top 1000 and applies the correct binary threshold.