| # 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. |
|
|
