Skip to content

Latest commit

 

History

History
110 lines (78 loc) · 6.99 KB

regressions-msmarco-v1-passage.bge-base-en-v1.5.parquet.hnsw-int8.onnx.md

File metadata and controls

110 lines (78 loc) · 6.99 KB
Raw

Anserini Regressions: MS MARCO Passage Ranking

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 the MS MARCO passage ranking task, 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 performing query inference "on-the-fly" with ONNX.

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 msmarco-v1-passage.bge-base-en-v1.5.parquet.hnsw-int8.onnx

We make available a version of the MS MARCO Passage Corpus that has already been encoded by the BGE-base-en-v1.5 model.

From any machine, the following command will download the corpus and perform the complete regression, end to end:

python src/main/python/run_regression.py --download --index --verify --search --regression msmarco-v1-passage.bge-base-en-v1.5.parquet.hnsw-int8.onnx

The run_regression.py script automates the following steps, but if you want to perform each step manually, simply copy/paste from the commands below and you'll obtain the same regression results.

Corpus Download

Download the corpus and unpack into collections/:

wget https://rgw.cs.uwaterloo.ca/pyserini/data/msmarco-passage-bge-base-en-v1.5.parquet.tar -P collections/
tar xvf collections/msmarco-passage-bge-base-en-v1.5.parquet.tar -C collections/

To confirm, msmarco-passage-bge-base-en-v1.5.parquet.tar is 39 GB and has MD5 checksum b235e19ec492c18a18057b30b8b23fd4. With the corpus downloaded, the following command will perform the remaining steps below:

python src/main/python/run_regression.py --index --verify --search --regression msmarco-v1-passage.bge-base-en-v1.5.parquet.hnsw-int8.onnx \
  --corpus-path collections/msmarco-passage-bge-base-en-v1.5.parquet

Indexing

Sample indexing command, building quantized HNSW indexes:

bin/run.sh io.anserini.index.IndexHnswDenseVectors \
  -threads 16 \
  -collection ParquetDenseVectorCollection \
  -input /path/to/msmarco-passage-bge-base-en-v1.5.parquet \
  -generator ParquetDenseVectorDocumentGenerator \
  -index indexes/lucene-hnsw-int8.msmarco-v1-passage.bge-base-en-v1.5/ \
  -M 16 -efC 100 -quantize.int8 \
  >& logs/log.msmarco-passage-bge-base-en-v1.5.parquet &

The path /path/to/msmarco-passage-bge-base-en-v1.5.parquet/ should point to the corpus downloaded above. Upon completion, we should have an index with 8,841,823 documents.

Retrieval

Topics and qrels are stored here, which is linked to the Anserini repo as a submodule. The regression experiments here evaluate on the 6980 dev set questions; see this page for more details.

After indexing has completed, you should be able to perform retrieval as follows using HNSW indexes:

bin/run.sh io.anserini.search.SearchHnswDenseVectors \
  -index indexes/lucene-hnsw-int8.msmarco-v1-passage.bge-base-en-v1.5/ \
  -topics tools/topics-and-qrels/topics.msmarco-passage.dev-subset.txt \
  -topicReader TsvInt \
  -output runs/run.msmarco-passage-bge-base-en-v1.5.parquet.bge-hnsw-int8-onnx.topics.msmarco-passage.dev-subset.txt \
  -encoder BgeBaseEn15 -hits 1000 -efSearch 1000 -threads 16 &

Evaluation can be performed using trec_eval:

bin/trec_eval -c -m map tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-bge-base-en-v1.5.parquet.bge-hnsw-int8-onnx.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -M 10 -m recip_rank tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-bge-base-en-v1.5.parquet.bge-hnsw-int8-onnx.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -m recall.100 tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-bge-base-en-v1.5.parquet.bge-hnsw-int8-onnx.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -m recall.1000 tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-bge-base-en-v1.5.parquet.bge-hnsw-int8-onnx.topics.msmarco-passage.dev-subset.txt

Effectiveness

With the above commands, you should be able to reproduce the following results:

AP@1000 BGE-base-en-v1.5
MS MARCO Passage: Dev 0.364
RR@10 BGE-base-en-v1.5
MS MARCO Passage: Dev 0.358
R@100 BGE-base-en-v1.5
MS MARCO Passage: Dev 0.901
R@1000 BGE-base-en-v1.5
MS MARCO Passage: Dev 0.981

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 are likely to differ; scores may be lower by up to 0.01, sometimes more. Note that both HNSW indexing and quantization are non-deterministic (i.e., results may differ slightly between trials).

Reproduction Log*

To add to this reproduction log, modify this template and run bin/build.sh to rebuild the documentation.