Skip to content

Latest commit

 

History

History
86 lines (62 loc) · 5.74 KB

regressions-beir-v1.0.0-cqadupstack-programmers.bge-base-en-v1.5.hnsw-int8.onnx.md

File metadata and controls

86 lines (62 loc) · 5.74 KB

Anserini Regressions: BEIR (v1.0.0) — CQADupStack-programmers

Model: BGE-base-en-v1.5 with quantized HNSW indexes (using ONNX for on-the-fly query encoding)

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-programmers, 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 ONNX to perform query encoding on the fly.

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-programmers.bge-base-en-v1.5.hnsw-int8.onnx

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.

Indexing

Sample indexing command, building quantized HNSW indexes:

bin/run.sh io.anserini.index.IndexHnswDenseVectors \
  -threads 16 \
  -collection JsonDenseVectorCollection \
  -input /path/to/beir-v1.0.0-cqadupstack-programmers.bge-base-en-v1.5 \
  -generator DenseVectorDocumentGenerator \
  -index indexes/lucene-hnsw-int8.beir-v1.0.0-cqadupstack-programmers.bge-base-en-v1.5/ \
  -M 16 -efC 100 -quantize.int8 \
  >& logs/log.beir-v1.0.0-cqadupstack-programmers.bge-base-en-v1.5 &

The path /path/to/beir-v1.0.0-cqadupstack-programmers.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.

Retrieval

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:

bin/run.sh io.anserini.search.SearchHnswDenseVectors \
  -index indexes/lucene-hnsw-int8.beir-v1.0.0-cqadupstack-programmers.bge-base-en-v1.5/ \
  -topics tools/topics-and-qrels/topics.beir-v1.0.0-cqadupstack-programmers.test.tsv.gz \
  -topicReader TsvString \
  -output runs/run.beir-v1.0.0-cqadupstack-programmers.bge-base-en-v1.5.bge-hnsw-int8-onnx.topics.beir-v1.0.0-cqadupstack-programmers.test.txt \
  -encoder BgeBaseEn15 -hits 1000 -efSearch 1000 -removeQuery -threads 16 &

Evaluation can be performed using trec_eval:

bin/trec_eval -c -m ndcg_cut.10 tools/topics-and-qrels/qrels.beir-v1.0.0-cqadupstack-programmers.test.txt runs/run.beir-v1.0.0-cqadupstack-programmers.bge-base-en-v1.5.bge-hnsw-int8-onnx.topics.beir-v1.0.0-cqadupstack-programmers.test.txt
bin/trec_eval -c -m recall.100 tools/topics-and-qrels/qrels.beir-v1.0.0-cqadupstack-programmers.test.txt runs/run.beir-v1.0.0-cqadupstack-programmers.bge-base-en-v1.5.bge-hnsw-int8-onnx.topics.beir-v1.0.0-cqadupstack-programmers.test.txt
bin/trec_eval -c -m recall.1000 tools/topics-and-qrels/qrels.beir-v1.0.0-cqadupstack-programmers.test.txt runs/run.beir-v1.0.0-cqadupstack-programmers.bge-base-en-v1.5.bge-hnsw-int8-onnx.topics.beir-v1.0.0-cqadupstack-programmers.test.txt

Effectiveness

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-programmers 0.424
R@100 BGE-base-en-v1.5
BEIR (v1.0.0): CQADupStack-programmers 0.786
R@1000 BGE-base-en-v1.5
BEIR (v1.0.0): CQADupStack-programmers 0.935

The above figures are from running brute-force search with cached queries on non-quantized flat indexes. With ONNX query encoding on quantized HNSW indexes, observed results may differ slightly (typically, lower), but scores should generally be within 0.005 of the results reported above (with some outliers). Note that both HNSW indexing and quantization are non-deterministic (i.e., results may differ slightly between trials).