Upload README.md with huggingface_hub
Browse files
README.md
ADDED
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Dataset for HybRank
|
2 |
+
|
3 |
+
You can download preprocessed data from [HuggingFace Repo](https://huggingface.co/datasets/ustc-zhangzm/HybRank)
|
4 |
+
|
5 |
+
Note that `train_scores.hdf5` of `MS MARCO` dataset files are split via
|
6 |
+
```bash
|
7 |
+
split -d -b 3G train_scores.hdf5 train_scores.hdf5.
|
8 |
+
```
|
9 |
+
|
10 |
+
Run following command to concatenate these files after all shards have been downloaded
|
11 |
+
```bash
|
12 |
+
cat train_scores.hdf5.* > train_scores.hdf5
|
13 |
+
```
|
14 |
+
|
15 |
+
Or you can generate data by yourself via the following steps:
|
16 |
+
|
17 |
+
## Dependencies
|
18 |
+
```
|
19 |
+
java 11.0.16
|
20 |
+
maven 3.8.6
|
21 |
+
anserini 0.14.3
|
22 |
+
faiss-cpu 1.7.2
|
23 |
+
pyserini 0.17.1
|
24 |
+
```
|
25 |
+
|
26 |
+
|
27 |
+
## Natural Questions
|
28 |
+
|
29 |
+
### 1. Download raw data (Refer to [DPR](https://github.com/facebookresearch/DPR) for more details of the dataset)
|
30 |
+
```shell
|
31 |
+
python download_DPR_data.py --resource data.wikipedia_split.psgs_w100
|
32 |
+
python download_DPR_data.py --resource data.retriever.nq
|
33 |
+
python download_DPR_data.py --resource data.retriever.qas.nq
|
34 |
+
mkdir -p raw && mv downloads raw/NQ
|
35 |
+
```
|
36 |
+
|
37 |
+
### 2. Convert collections to jsonl format for Pyserini
|
38 |
+
```shell
|
39 |
+
python convert_NQ_collection_to_jsonl.py --collection-path raw/NQ/data/wikipedia_split/psgs_w100.tsv --output-folder pyserini/collections/NQ
|
40 |
+
```
|
41 |
+
|
42 |
+
### 3. Build Lucene indexes via Pyserini
|
43 |
+
```shell
|
44 |
+
python -m pyserini.index.lucene \
|
45 |
+
--collection JsonCollection \
|
46 |
+
--input pyserini/collections/NQ \
|
47 |
+
--index pyserini/indexes/NQ \
|
48 |
+
--generator DefaultLuceneDocumentGenerator \
|
49 |
+
--threads 1 \
|
50 |
+
--storePositions --storeDocvectors --storeRaw
|
51 |
+
```
|
52 |
+
|
53 |
+
### 4. Generate data
|
54 |
+
```shell
|
55 |
+
RETRIEVERS=("DPR-Multi" "DPR-Single" "ANCE" "FiD-KD" "RocketQA-retriever" "RocketQAv2-retriever" "RocketQA-reranker" "RocketQAv2-reranker")
|
56 |
+
|
57 |
+
for RETRIEVER in ${RETRIEVERS[@]}; do
|
58 |
+
python generate_NQ_data.py --retriever $RETRIEVER
|
59 |
+
done
|
60 |
+
```
|
61 |
+
Note that before generate data for retriever `RocketQA*`, please generate the retrieval results following the instructions in `data/RocketQA_baselines/README.md`. Data for other retrievers can be generated directly.
|
62 |
+
|
63 |
+
|
64 |
+
## MS MARCO & TREC 2019/2020
|
65 |
+
|
66 |
+
### 1. Download raw data (Refer to [MS MARCO](https://microsoft.github.io/msmarco/) for more details of the dataset)
|
67 |
+
* Download and uncompress MSMARCO Passage Ranking Collections and Queries [collectionandqueries.tar.gz](https://msmarco.blob.core.windows.net/msmarcoranking/collectionandqueries.tar.gz) to `data/raw/MSMARCO/`
|
68 |
+
* TREC DL Test Queries and Qrels
|
69 |
+
* [TREC DL-2019](https://microsoft.github.io/msmarco/TREC-Deep-Learning-2019.html)
|
70 |
+
* Download and uncompress [msmarco-test2019-queries.tsv](https://msmarco.blob.core.windows.net/msmarcoranking/msmarco-test2019-queries.tsv.gz)
|
71 |
+
* Download [2019qrels-pass.txt](https://trec.nist.gov/data/deep/2019qrels-pass.txt)
|
72 |
+
* [TREC DL-2020](https://microsoft.github.io/msmarco/TREC-Deep-Learning-2020)
|
73 |
+
* Download and uncompress [msmarco-test2019-queries.tsv](https://msmarco.blob.core.windows.net/msmarcoranking/msmarco-test2020-queries.tsv.gz)
|
74 |
+
* Download [2019qrels-pass.txt](https://trec.nist.gov/data/deep/2020qrels-pass.txt)
|
75 |
+
* Put them into `data/raw/TRECDL/`
|
76 |
+
|
77 |
+
### 2. Convert collections to jsonl format for Pyserini
|
78 |
+
|
79 |
+
```shell
|
80 |
+
python convert_MSMARCO_collection_to_jsonl.py --collection-path raw/MSMARCO/collection.tsv --output-folder pyserini/collections/MSMARCO
|
81 |
+
```
|
82 |
+
|
83 |
+
### 3. Build Lucene indexes via Pyserini
|
84 |
+
|
85 |
+
```shell
|
86 |
+
python -m pyserini.index.lucene \
|
87 |
+
--collection JsonCollection \
|
88 |
+
--input pyserini/collections/MSMARCO \
|
89 |
+
--index pyserini/indexes/MSMARCO \
|
90 |
+
--generator DefaultLuceneDocumentGenerator \
|
91 |
+
--threads 1 \
|
92 |
+
--storePositions --storeDocvectors --storeRaw
|
93 |
+
```
|
94 |
+
|
95 |
+
### 4. Generate data
|
96 |
+
```shell
|
97 |
+
RETRIEVERS=("ANCE" "DistilBERT-KD" "TAS-B" "TCT-ColBERT-v1" "TCT-ColBERT-v2" "RocketQA-retriever" "RocketQAv2-retriever" "RocketQA-reranker" "RocketQAv2-reranker")
|
98 |
+
|
99 |
+
for RETRIEVER in ${RETRIEVERS[@]}; do
|
100 |
+
python generate_MSMARCO_data.py --retriever $RETRIEVER
|
101 |
+
done
|
102 |
+
```
|
103 |
+
|
104 |
+
```shell
|
105 |
+
RETRIEVERS=("ANCE" "DistilBERT-KD" "TAS-B" "TCT-ColBERT-v1" "TCT-ColBERT-v2" "RocketQA-retriever" "RocketQAv2-retriever" "RocketQA-reranker" "RocketQAv2-reranker")
|
106 |
+
|
107 |
+
SPLITS=("2019" "2020")
|
108 |
+
|
109 |
+
for RETRIEVER in ${RETRIEVERS[@]}; do
|
110 |
+
for SPLIT in ${SPLITS[@]}; do
|
111 |
+
python generate_TRECDL_data.py --split $SPLIT --retriever $RETRIEVER
|
112 |
+
done
|
113 |
+
done
|
114 |
+
```
|