MS MARCO Subset
Purpose
This subset provides a standardized benchmark for evaluating sparse model performance on MS MARCO data, with negative examples pooled using BM25 retrieval.
Creation
This subset was created using the make_beir_subset.py script with the following command:
python ./make_beir_subset.py --dataset msmarco --es-host http://localhost:9200 --force-reindex --split dev
Parameters Used
- Dataset: msmarco
- Split: dev (development set)
- ES Host: http://localhost:9200
- Force Reindex: Enabled (overwrote existing Elasticsearch index)
- Default settings:
- Number of queries: 1000 (or available queries if fewer)
- Pool size: 100 documents per query
- Pooler: elastic (BM25 retrieval)
- Seed: 42