## Introduction

This respository introduces how to reproduce the `Dense`, `Sparse`, and `Dense+Sparse` evaluation results of the paper [BGE-M3](https://arxiv.org/pdf/2402.03216.pdf) on the [MIRACL](https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00595/117438/MIRACL-A-Multilingual-Retrieval-Dataset-Covering) dev split.

## Requirements

```bash
# Install Java (Linux)
apt update
apt install openjdk-21-jdk

# Install Pyserini
pip install pyserini

# Install Faiss
## CPU version
conda install -c conda-forge faiss-cpu

## GPU version
conda install -c conda-forge faiss-gpu
```

**It should be noted that** the Pyserini code needs to be modified to support the multiple alpha settings in `pyserini/fusion`. I have already submitted a pull request to the official repository to support this feature. You can refer to this [PR](https://github.com/castorini/pyserini/pull/1858) to modify the code.

## 2CR

### Download and Unzip

```bash
# Download
## MIRACL topics and qrels
git clone https://huggingface.co/datasets/miracl/miracl
mv miracl/*/*/* topics-and-qrels
## Dense and Sparse Index
git lfs install
git clone https://huggingface.co/datasets/hanhainebula/bge-m3_miracl_2cr

cat bge-m3_miracl_2cr/dense/en.tar.gz.part_* > bge-m3_miracl_2cr/dense/en.tar.gz
cat bge-m3_miracl_2cr/dense/de.tar.gz.part_* > bge-m3_miracl_2cr/dense/de.tar.gz


# Unzip
languages=(ar bn en es fa fi fr hi id ja ko ru sw te th zh de yo)

## Dense
for lang in ${languages[@]}; do
  tar -zxvf bge-m3_miracl_2cr/dense/${lang}.tar.gz -C bge-m3_miracl_2cr/dense/
done

## Sparse
for lang in ${languages[@]}; do
  tar -zxvf bge-m3_miracl_2cr/sparse/${lang}.tar.gz -C bge-m3_miracl_2cr/sparse/
done
```

### Reproduction

#### Dense

```bash
# Avaliable Language: ar bn en es fa fi fr hi id ja ko ru sw te th zh de yo
lang=zh

# Generate run
python -m pyserini.search.faiss \
  --threads 16 --batch-size 512 \
  --encoder-class auto \
  --encoder BAAI/bge-m3 \
  --pooling cls --l2-norm \
  --topics topics-and-qrels/topics.miracl-v1.0-${lang}-dev.tsv \
  --index bge-m3_miracl_2cr/dense/${lang} \
  --output bge-m3_miracl_2cr/dense/runs/${lang}.txt \
  --hits 1000

# Evaluate
## nDCG@10
python -m pyserini.eval.trec_eval \
  -c -M 100 -m ndcg_cut.10 \
  topics-and-qrels/qrels.miracl-v1.0-${lang}-dev.tsv \
  bge-m3_miracl_2cr/dense/runs/${lang}.txt
## Recall@100
python -m pyserini.eval.trec_eval \
  -c -m recall.100 \
  topics-and-qrels/qrels.miracl-v1.0-${lang}-dev.tsv \
  bge-m3_miracl_2cr/dense/runs/${lang}.txt
```

#### Sparse

```bash
# Avaliable Language: ar bn en es fa fi fr hi id ja ko ru sw te th zh de yo
lang=zh

# Generate run
python -m pyserini.search.lucene \
  --threads 16 --batch-size 128 \
  --topics bge-m3_miracl_2cr/sparse/${lang}/query_embd.tsv \
  --index bge-m3_miracl_2cr/sparse/${lang}/index \
  --output bge-m3_miracl_2cr/sparse/runs/${lang}.txt \
  --output-format trec \
  --impact --hits 1000

# Evaluate
## nDCG@10
python -m pyserini.eval.trec_eval \
  -c -M 100 -m ndcg_cut.10 \
  topics-and-qrels/qrels.miracl-v1.0-${lang}-dev.tsv \
  bge-m3_miracl_2cr/sparse/runs/${lang}.txt
## Recall@100
python -m pyserini.eval.trec_eval \
  -c -m recall.100 \
  topics-and-qrels/qrels.miracl-v1.0-${lang}-dev.tsv \
  bge-m3_miracl_2cr/sparse/runs/${lang}.txt
```

#### Dense+Sparse

**Note**: You should first merge this [PR](https://github.com/castorini/pyserini/pull/1858) to support the multiple alpha settings in `pyserini/fusion`.

```bash
# Avaliable Language: ar bn en es fa fi fr hi id ja ko ru sw te th zh de yo
lang=zh

# Generate dense run and sparse run
python -m pyserini.search.faiss \
  --threads 16 --batch-size 512 \
  --encoder-class auto \
  --encoder BAAI/bge-m3 \
  --pooling cls --l2-norm \
  --topics topics-and-qrels/topics.miracl-v1.0-${lang}-dev.tsv \
  --index bge-m3_miracl_2cr/dense/${lang} \
  --output bge-m3_miracl_2cr/dense/runs/${lang}.txt \
  --hits 1000

python -m pyserini.search.lucene \
  --threads 16 --batch-size 128 \
  --topics bge-m3_miracl_2cr/sparse/${lang}/query_embd.tsv \
  --index bge-m3_miracl_2cr/sparse/${lang}/index \
  --output bge-m3_miracl_2cr/sparse/runs/${lang}.txt \
  --output-format trec \
  --impact --hits 1000

# Generate dense+sparse run
mkdir -p bge-m3_miracl_2cr/fusion/runs

python -m pyserini.fusion \
  --method interpolation \
  --runs bge-m3_miracl_2cr/dense/runs/${lang}.txt bge-m3_miracl_2cr/sparse/runs/${lang}.txt \
  --alpha 1 3e-5 \
  --output bge-m3_miracl_2cr/fusion/runs/${lang}.txt \
  --depth 1000 --k 1000

# Evaluation
## nDCG@10
python -m pyserini.eval.trec_eval \
  -c -M 100 -m ndcg_cut.10 \
  topics-and-qrels/qrels.miracl-v1.0-${lang}-dev.tsv \
  bge-m3_miracl_2cr/fusion/runs/${lang}.txt
## Recall@100
python -m pyserini.eval.trec_eval \
  -c -m recall.100 \
  topics-and-qrels/qrels.miracl-v1.0-${lang}-dev.tsv \
  bge-m3_miracl_2cr/fusion/runs/${lang}.txt
```

Note:
- The hybrid method we used for MIRACL in BGE-M3 paper is: `s_dense + 0.3 * s_sparse`. But when the sparse score is calculated, it has already been multiplied by 100^2, so the alpha for sparse run here is 3e-5, instead of 0.3.