Model: BGE-base-en-v1.5 with quantized HNSW indexes (using pre-encoded queries)
This page describes regression experiments, integrated into Anserini's regression testing framework, using the BGE-base-en-v1.5 model on BEIR (v1.0.0) — CQADupStack-stats, as described in the following paper:
Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv:2309.07597, 2023.
In these experiments, we are using pre-encoded queries (i.e., cached results of query encoding).
The exact configurations for these regressions are stored in this YAML file.
Note that this page is automatically generated from this template as part of Anserini's regression pipeline, so do not modify this page directly; modify the template instead and then run bin/build.sh to rebuild the documentation.
From one of our Waterloo servers (e.g., orca), the following command will perform the complete regression, end to end:
python src/main/python/run_regression.py --index --verify --search --regression beir-v1.0.0-cqadupstack-stats-bge-base-en-v1.5-hnsw-int8
All the BEIR corpora, encoded by the BGE-base-en-v1.5 model, are available for download:
wget https://rgw.cs.uwaterloo.ca/pyserini/data/beir-v1.0.0-bge-base-en-v1.5.tar -P collections/
tar xvf collections/beir-v1.0.0-bge-base-en-v1.5.tar -C collections/The tarball is 294 GB and has MD5 checksum e4e8324ba3da3b46e715297407a24f00.
After download and unpacking the corpora, the run_regression.py command above should work without any issue.
Sample indexing command, building HNSW indexes:
target/appassembler/bin/IndexHnswDenseVectors \
-collection JsonDenseVectorCollection \
-input /path/to/beir-v1.0.0-bge-base-en-v1.5 \
-generator HnswDenseVectorDocumentGenerator \
-index indexes/lucene-hnsw.beir-v1.0.0-cqadupstack-stats-bge-base-en-v1.5-int8/ \
-threads 16 -M 16 -efC 100 -memoryBuffer 65536 -noMerge -quantize.int8 \
>& logs/log.beir-v1.0.0-bge-base-en-v1.5 &
The path /path/to/beir-v1.0.0-bge-base-en-v1.5/ should point to the corpus downloaded above.
Note that here we are explicitly using Lucene's NoMergePolicy merge policy, which suppresses any merging of index segments.
This is because merging index segments is a costly operation and not worthwhile given our query set.
Topics and qrels are stored here, which is linked to the Anserini repo as a submodule.
After indexing has completed, you should be able to perform retrieval as follows:
target/appassembler/bin/SearchHnswDenseVectors \
-index indexes/lucene-hnsw.beir-v1.0.0-cqadupstack-stats-bge-base-en-v1.5-int8/ \
-topics tools/topics-and-qrels/topics.beir-v1.0.0-cqadupstack-stats.test.bge-base-en-v1.5.jsonl.gz \
-topicReader JsonStringVector \
-output runs/run.beir-v1.0.0-bge-base-en-v1.5.bge-hnsw.topics.beir-v1.0.0-cqadupstack-stats.test.bge-base-en-v1.5.jsonl.txt \
-generator VectorQueryGenerator -topicField vector -removeQuery -threads 16 -hits 1000 -efSearch 1000 &
Evaluation can be performed using trec_eval:
target/appassembler/bin/trec_eval -c -m ndcg_cut.10 tools/topics-and-qrels/qrels.beir-v1.0.0-cqadupstack-stats.test.txt runs/run.beir-v1.0.0-bge-base-en-v1.5.bge-hnsw.topics.beir-v1.0.0-cqadupstack-stats.test.bge-base-en-v1.5.jsonl.txt
target/appassembler/bin/trec_eval -c -m recall.100 tools/topics-and-qrels/qrels.beir-v1.0.0-cqadupstack-stats.test.txt runs/run.beir-v1.0.0-bge-base-en-v1.5.bge-hnsw.topics.beir-v1.0.0-cqadupstack-stats.test.bge-base-en-v1.5.jsonl.txt
target/appassembler/bin/trec_eval -c -m recall.1000 tools/topics-and-qrels/qrels.beir-v1.0.0-cqadupstack-stats.test.txt runs/run.beir-v1.0.0-bge-base-en-v1.5.bge-hnsw.topics.beir-v1.0.0-cqadupstack-stats.test.bge-base-en-v1.5.jsonl.txt
With the above commands, you should be able to reproduce the following results:
| nDCG@10 | BGE-base-en-v1.5 |
|---|---|
| BEIR (v1.0.0): CQADupStack-stats | 0.371 |
| R@100 | BGE-base-en-v1.5 |
| BEIR (v1.0.0): CQADupStack-stats | 0.672 |
| R@1000 | BGE-base-en-v1.5 |
| BEIR (v1.0.0): CQADupStack-stats | 0.852 |
Note that due to the non-deterministic nature of HNSW indexing, results may differ slightly between each experimental run. Nevertheless, scores are generally within 0.005 of the reference values recorded in our YAML configuration file.