Update README.md
Browse files
README.md
CHANGED
@@ -42,21 +42,52 @@ configs:
|
|
42 |
path: "qrels/test/*.txt"
|
43 |
---
|
44 |
|
45 |
-
|
|
|
46 |
|
47 |
-
Project Page: [https://tiger-ai-lab.github.io/UniIR/](https://tiger-ai-lab.github.io/UniIR/)
|
48 |
|
49 |
-
|
50 |
|
51 |
-
|
|
|
|
|
52 |
|
53 |
|
54 |
## **Dataset Structure Overview**
|
55 |
-
M-BEIR dataset
|
56 |
-
Each of these sections consists of structured entries in JSONL format (JSON Lines), meaning each line is a valid JSON object. Below is a detailed breakdown of the components and their respective fields:
|
57 |
|
58 |
-
Query Data
|
59 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
```json
|
61 |
{
|
62 |
"qid": "A unique identifier formatted as {dataset_id}:{query_id}",
|
@@ -79,8 +110,30 @@ Each line in the Query Data file represents a unique query. The structure of eac
|
|
79 |
}
|
80 |
```
|
81 |
|
82 |
-
Candidate Pool
|
83 |
-
The Candidate Pool contains potential matching documents for the queries.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
```json
|
85 |
{
|
86 |
"did": "A unique identifier for the document, formatted as {dataset_id}:{doc_id}",
|
@@ -91,9 +144,16 @@ The Candidate Pool contains potential matching documents for the queries. The st
|
|
91 |
}
|
92 |
```
|
93 |
|
|
|
|
|
|
|
|
|
|
|
|
|
94 |
## **How to Use**
|
95 |
### Downloading the M-BEIR Dataset
|
96 |
-
|
|
|
97 |
|
98 |
### Decompressing M-BEIR Images
|
99 |
After downloading, you will need to decompress the image files. Follow these steps in your terminal:
|
@@ -107,9 +167,12 @@ sh -c 'cat mbeir_images.tar.gz.part-00 mbeir_images.tar.gz.part-01 mbeir_images.
|
|
107 |
# Extract the images from the tar.gz file
|
108 |
tar -xzf mbeir_images.tar.gz
|
109 |
```
|
|
|
110 |
|
111 |
-
|
|
|
112 |
|
|
|
113 |
Please cite our paper if you use our data, model or code.
|
114 |
|
115 |
```
|
|
|
42 |
path: "qrels/test/*.txt"
|
43 |
---
|
44 |
|
45 |
+
### **UniIR: Training and Benchmarking Universal Multimodal Information Retrievers**
|
46 |
+
[**π Homepage**](https://tiger-ai-lab.github.io/UniIR/) | [**π€ Paper**](https://huggingface.co/papers/2311.17136) | [**π arXiv**](https://arxiv.org/pdf/2311.17136.pdf) | [**GitHub**](https://github.com/TIGER-AI-Lab/UniIR)
|
47 |
|
|
|
48 |
|
49 |
+
## **Dataset Summary**
|
50 |
|
51 |
+
M-BEIR, the **M**ultimodal **BE**nchmark for **I**nstructed **R**etrieval, is a comprehensive large-scale retrieval benchmark designed to train and evaluate unified multimodal retrieval models (**UniIR models**).
|
52 |
+
The M-BEIR benchmark comprises eight multimodal retrieval tasks and ten datasets from a variety of domains and sources.
|
53 |
+
Each task is accompanied by human-authored instructions, encompassing 1.5 million queries and a pool of 5.6 million retrieval candidates in total.
|
54 |
|
55 |
|
56 |
## **Dataset Structure Overview**
|
57 |
+
The M-BEIR dataset is structured into five primary components: Query Data, Candidate Pool, Instructions, Qrels, and Images.
|
|
|
58 |
|
59 |
+
### Query Data
|
60 |
+
|
61 |
+
Below is the directory structure for the query data:
|
62 |
+
```
|
63 |
+
query/
|
64 |
+
β
|
65 |
+
βββ train/
|
66 |
+
β βββ mbeir_cirr_train.jsonl
|
67 |
+
β βββ mbeir_edis_train.jsonl
|
68 |
+
β ...
|
69 |
+
βββ union_train/
|
70 |
+
β βββ mbeir_union_up_train.jsonl
|
71 |
+
βββ val/
|
72 |
+
β βββ mbeir_visualnews_task0_val.jsonl
|
73 |
+
β βββ mbeir_visualnews_task3_val.jsonl
|
74 |
+
β ...
|
75 |
+
βββ test/
|
76 |
+
βββ mbeir_visualnews_task0_test.jsonl
|
77 |
+
βββ mbeir_visualnews_task3_test.jsonl
|
78 |
+
...
|
79 |
+
```
|
80 |
+
`train`: Contains all the training data from 8 different datasets formatted in the M-BEIR style.
|
81 |
+
|
82 |
+
`mbeir_union_up_train.jsonl`: This file is the default training data for in-batch contrastive training specifically designed for UniIR models.
|
83 |
+
It aggregates all the data from the train directory and datasets with relatively smaller sizes have been upsampled to balance the training process.
|
84 |
+
|
85 |
+
`val`: Contains separate files for validation queries, organized by task.
|
86 |
+
|
87 |
+
`test`: Contains separate files for test queries, organized by task.
|
88 |
+
|
89 |
+
Every M-BEIR query instance has at least one positive candidate data and possibly no negative candidate data
|
90 |
+
Each line in a Query Data file represents a unique query. The structure of each query JSON object is as follows::
|
91 |
```json
|
92 |
{
|
93 |
"qid": "A unique identifier formatted as {dataset_id}:{query_id}",
|
|
|
110 |
}
|
111 |
```
|
112 |
|
113 |
+
### Candidate Pool
|
114 |
+
The Candidate Pool contains potential matching documents for the queries.
|
115 |
+
|
116 |
+
#### M-BEIR_5.6M
|
117 |
+
Within the global directory, the default retrieval setting requires models to retrieve positive candidates from a heterogeneous pool encompassing various modalities and domains.
|
118 |
+
The M-BEIR's global candidate pool, comprising 5.6 million candidates, includes the retrieval corpus from all tasks and datasets.
|
119 |
+
|
120 |
+
#### M-BEIR_local
|
121 |
+
Within the local directory, we provide dataset-task-specific pool as M-BEIR_local. Dataset-task-specific pool contains homogeneous candidates that originate from by the original dataset.
|
122 |
+
|
123 |
+
Below is the directory structure for the candidate pool:
|
124 |
+
```
|
125 |
+
cand_pool/
|
126 |
+
β
|
127 |
+
βββ global/
|
128 |
+
β βββ mbeir_union_val_cand_pool.jsonl
|
129 |
+
β βββmbeir_union_test_cand_pool.jsonl
|
130 |
+
β
|
131 |
+
βββ local/
|
132 |
+
βββ mbeir_visualnews_task0_cand_pool.jsonl
|
133 |
+
βββ mbeir_visualnews_task3_cand_pool.jsonl
|
134 |
+
...
|
135 |
+
```
|
136 |
+
The structure of each candidate JSON object in cand_pool file is as follows::
|
137 |
```json
|
138 |
{
|
139 |
"did": "A unique identifier for the document, formatted as {dataset_id}:{doc_id}",
|
|
|
144 |
}
|
145 |
```
|
146 |
|
147 |
+
### Instructions
|
148 |
+
`query_instructions.tsv` contains human-authorized instructions within the UniIR framework. Each task is accompanied by four human-authored instructions.
|
149 |
+
|
150 |
+
### Qrels
|
151 |
+
Within the `qrels` directory, you will find qrels for both the validation and test sets. These files serve the purpose of evaluating UniIR models.
|
152 |
+
|
153 |
## **How to Use**
|
154 |
### Downloading the M-BEIR Dataset
|
155 |
+
Clone the M-BEIR repo from the current Page.
|
156 |
+
Ensure that Git LFS (Large File Storage) is installed on your system, as it will download the required data files.
|
157 |
|
158 |
### Decompressing M-BEIR Images
|
159 |
After downloading, you will need to decompress the image files. Follow these steps in your terminal:
|
|
|
167 |
# Extract the images from the tar.gz file
|
168 |
tar -xzf mbeir_images.tar.gz
|
169 |
```
|
170 |
+
Now, you are ready to use the M-BEIR benchmark.
|
171 |
|
172 |
+
### Dataloader and Evaluation Pipeline
|
173 |
+
We offer a dedicated dataloader and evaluation pipeline for the M-BEIR benchmark. Please refer to [**GitHub Repo**](https://github.com/TIGER-AI-Lab/UniIR) for detailed information.
|
174 |
|
175 |
+
## **Citation**
|
176 |
Please cite our paper if you use our data, model or code.
|
177 |
|
178 |
```
|