Hugging Face's logo

Datasets:
nvidia
/
miracl-vision

Modalities:
Image
Tabular
Text
Formats:
parquet
Sub-tasks:
document-retrieval
Languages:
Arabic
Bengali
English
Size:
100K - 1M
ArXiv:
Tags:
text
image
Libraries:
Datasets
pandas
License:
Dataset card Data Studio Files
xet
Community
1
nv-bschifferer commited on
Commit
1bd8bbc
·
1 Parent(s): 7329a2e

README

Browse files
Files changed (1) hide show
  1. README.md +470 -0
README.md ADDED
@@ -0,0 +1,470 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - ar
4
+ - bn
5
+ - en
6
+ - es
7
+ - fa
8
+ - fi
9
+ - fr
10
+ - hi
11
+ - id
12
+ - ja
13
+ - ko
14
+ - ru
15
+ - sw
16
+ - te
17
+ - th
18
+ - zh
19
+ - de
20
+ - yo
21
+ multilinguality:
22
+ - multilingual
23
+ license:
24
+ - cc-by-sa-4.0
25
+ task_ids:
26
+ - document-retrieval
27
+ tags:
28
+ - image
29
+ configs:
30
+ - config_name: queries-ar
31
+ data_files:
32
+ - split: default
33
+ path: "ar/queries.parquet"
34
+ - config_name: corpus-ar
35
+ data_files:
36
+ - split: default
37
+ path: "ar/corpus.parquet"
38
+ - config_name: qrels-ar
39
+ data_files:
40
+ - split: default
41
+ path: "ar/qrels.parquet"
42
+ - config_name: images-ar
43
+ data_files:
44
+ - split: default
45
+ path: "ar/images.parquet"
46
+ - config_name: queries-bn
47
+ data_files:
48
+ - split: default
49
+ path: "bn/queries.parquet"
50
+ - config_name: corpus-bn
51
+ data_files:
52
+ - split: default
53
+ path: "bn/corpus.parquet"
54
+ - config_name: qrels-bn
55
+ data_files:
56
+ - split: default
57
+ path: "bn/qrels.parquet"
58
+ - config_name: images-bn
59
+ data_files:
60
+ - split: default
61
+ path: "bn/images.parquet"
62
+ - config_name: queries-de
63
+ data_files:
64
+ - split: default
65
+ path: "de/queries.parquet"
66
+ - config_name: corpus-de
67
+ data_files:
68
+ - split: default
69
+ path: "de/corpus.parquet"
70
+ - config_name: qrels-de
71
+ data_files:
72
+ - split: default
73
+ path: "de/qrels.parquet"
74
+ - config_name: images-de
75
+ data_files:
76
+ - split: default
77
+ path: "de/images.parquet"
78
+ - config_name: queries-en
79
+ data_files:
80
+ - split: default
81
+ path: "en/queries.parquet"
82
+ - config_name: corpus-en
83
+ data_files:
84
+ - split: default
85
+ path: "en/corpus.parquet"
86
+ - config_name: qrels-en
87
+ data_files:
88
+ - split: default
89
+ path: "en/qrels.parquet"
90
+ - config_name: images-en
91
+ data_files:
92
+ - split: default
93
+ path: "en/images.parquet"
94
+ - config_name: queries-es
95
+ data_files:
96
+ - split: default
97
+ path: "es/queries.parquet"
98
+ - config_name: corpus-es
99
+ data_files:
100
+ - split: default
101
+ path: "es/corpus.parquet"
102
+ - config_name: qrels-es
103
+ data_files:
104
+ - split: default
105
+ path: "es/qrels.parquet"
106
+ - config_name: images-es
107
+ data_files:
108
+ - split: default
109
+ path: "es/images.parquet"
110
+ - config_name: queries-fa
111
+ data_files:
112
+ - split: default
113
+ path: "fa/queries.parquet"
114
+ - config_name: corpus-fa
115
+ data_files:
116
+ - split: default
117
+ path: "fa/corpus.parquet"
118
+ - config_name: qrels-fa
119
+ data_files:
120
+ - split: default
121
+ path: "fa/qrels.parquet"
122
+ - config_name: images-fa
123
+ data_files:
124
+ - split: default
125
+ path: "fa/images.parquet"
126
+ - config_name: queries-fi
127
+ data_files:
128
+ - split: default
129
+ path: "fi/queries.parquet"
130
+ - config_name: corpus-fi
131
+ data_files:
132
+ - split: default
133
+ path: "fi/corpus.parquet"
134
+ - config_name: qrels-fi
135
+ data_files:
136
+ - split: default
137
+ path: "fi/qrels.parquet"
138
+ - config_name: images-fi
139
+ data_files:
140
+ - split: default
141
+ path: "fi/images.parquet"
142
+ - config_name: queries-fr
143
+ data_files:
144
+ - split: default
145
+ path: "fr/queries.parquet"
146
+ - config_name: corpus-fr
147
+ data_files:
148
+ - split: default
149
+ path: "fr/corpus.parquet"
150
+ - config_name: qrels-fr
151
+ data_files:
152
+ - split: default
153
+ path: "fr/qrels.parquet"
154
+ - config_name: images-fr
155
+ data_files:
156
+ - split: default
157
+ path: "fr/images.parquet"
158
+ - config_name: queries-hi
159
+ data_files:
160
+ - split: default
161
+ path: "hi/queries.parquet"
162
+ - config_name: corpus-hi
163
+ data_files:
164
+ - split: default
165
+ path: "hi/corpus.parquet"
166
+ - config_name: qrels-hi
167
+ data_files:
168
+ - split: default
169
+ path: "hi/qrels.parquet"
170
+ - config_name: images-hi
171
+ data_files:
172
+ - split: default
173
+ path: "hi/images.parquet"
174
+ - config_name: queries-id
175
+ data_files:
176
+ - split: default
177
+ path: "id/queries.parquet"
178
+ - config_name: corpus-id
179
+ data_files:
180
+ - split: default
181
+ path: "id/corpus.parquet"
182
+ - config_name: qrels-id
183
+ data_files:
184
+ - split: default
185
+ path: "id/qrels.parquet"
186
+ - config_name: images-id
187
+ data_files:
188
+ - split: default
189
+ path: "id/images.parquet"
190
+ - config_name: queries-ja
191
+ data_files:
192
+ - split: default
193
+ path: "ja/queries.parquet"
194
+ - config_name: corpus-ja
195
+ data_files:
196
+ - split: default
197
+ path: "ja/corpus.parquet"
198
+ - config_name: qrels-ja
199
+ data_files:
200
+ - split: default
201
+ path: "ja/qrels.parquet"
202
+ - config_name: images-ja
203
+ data_files:
204
+ - split: default
205
+ path: "ja/images.parquet"
206
+ - config_name: queries-ko
207
+ data_files:
208
+ - split: default
209
+ path: "ko/queries.parquet"
210
+ - config_name: corpus-ko
211
+ data_files:
212
+ - split: default
213
+ path: "ko/corpus.parquet"
214
+ - config_name: qrels-ko
215
+ data_files:
216
+ - split: default
217
+ path: "ko/qrels.parquet"
218
+ - config_name: images-ko
219
+ data_files:
220
+ - split: default
221
+ path: "ko/images.parquet"
222
+ - config_name: queries-ru
223
+ data_files:
224
+ - split: default
225
+ path: "ru/queries.parquet"
226
+ - config_name: corpus-ru
227
+ data_files:
228
+ - split: default
229
+ path: "ru/corpus.parquet"
230
+ - config_name: qrels-ru
231
+ data_files:
232
+ - split: default
233
+ path: "ru/qrels.parquet"
234
+ - config_name: images-ru
235
+ data_files:
236
+ - split: default
237
+ path: "ru/images.parquet"
238
+ - config_name: queries-sw
239
+ data_files:
240
+ - split: default
241
+ path: "sw/queries.parquet"
242
+ - config_name: corpus-sw
243
+ data_files:
244
+ - split: default
245
+ path: "sw/corpus.parquet"
246
+ - config_name: qrels-sw
247
+ data_files:
248
+ - split: default
249
+ path: "sw/qrels.parquet"
250
+ - config_name: images-sw
251
+ data_files:
252
+ - split: default
253
+ path: "sw/images.parquet"
254
+ - config_name: queries-te
255
+ data_files:
256
+ - split: default
257
+ path: "te/queries.parquet"
258
+ - config_name: corpus-te
259
+ data_files:
260
+ - split: default
261
+ path: "te/corpus.parquet"
262
+ - config_name: qrels-te
263
+ data_files:
264
+ - split: default
265
+ path: "te/qrels.parquet"
266
+ - config_name: images-te
267
+ data_files:
268
+ - split: default
269
+ path: "te/images.parquet"
270
+ - config_name: queries-th
271
+ data_files:
272
+ - split: default
273
+ path: "th/queries.parquet"
274
+ - config_name: corpus-th
275
+ data_files:
276
+ - split: default
277
+ path: "th/corpus.parquet"
278
+ - config_name: qrels-th
279
+ data_files:
280
+ - split: default
281
+ path: "th/qrels.parquet"
282
+ - config_name: images-th
283
+ data_files:
284
+ - split: default
285
+ path: "th/images.parquet"
286
+ - config_name: queries-yo
287
+ data_files:
288
+ - split: default
289
+ path: "yo/queries.parquet"
290
+ - config_name: corpus-yo
291
+ data_files:
292
+ - split: default
293
+ path: "yo/corpus.parquet"
294
+ - config_name: qrels-yo
295
+ data_files:
296
+ - split: default
297
+ path: "yo/qrels.parquet"
298
+ - config_name: images-yo
299
+ data_files:
300
+ - split: default
301
+ path: "yo/images.parquet"
302
+ - config_name: queries-zh
303
+ data_files:
304
+ - split: default
305
+ path: "zh/queries.parquet"
306
+ - config_name: corpus-zh
307
+ data_files:
308
+ - split: default
309
+ path: "zh/corpus.parquet"
310
+ - config_name: qrels-zh
311
+ data_files:
312
+ - split: default
313
+ path: "zh/qrels.parquet"
314
+ - config_name: images-zh
315
+ data_files:
316
+ - split: default
317
+ path: "zh/images.parquet"
318
+ ---
319
+
320
+ # MIRACL-VISION
321
+
322
+ MIRACL-VISION is a multilingual visual retrieval dataset for 18 different languages. It is an extension of MIRACL, a popular text-only multilingual retrieval dataset. The dataset contains user questions, images of Wikipedia articles and annotations, which article can answer a user question. There are 7898 questions and 338734 images.
323
+
324
+ This dataset is ready for commercial.
325
+
326
+ ### Correspondence to
327
+ Benedikt Schifferer ([email protected])
328
+
329
+ ### Dataset Creation Date:
330
+ 31st January 2025
331
+
332
+ ### License/Terms of Use:
333
+ This dataset is licensed under Creative Commons Attribution-ShareAlike 4.0 International. Additional Information: Apache License 2.0.
334
+
335
+ ### Intended Usage:
336
+ Users can evaluate multilingual, multimodal retriever pipelines.
337
+
338
+ ### Dataset Characterization
339
+ Dataset Collection Method: Automated
340
+ Labelling Method: Human
341
+
342
+ ### Dataset Format
343
+ The images are stored Pillow (PIL) Images in HuggingFace Dataset format
344
+ The questions, corpus, questions-corpus pairs are stored in parquet/BEIR format
345
+
346
+ ### Reference(s):
347
+ Paper to be published.
348
+
349
+ ### Ethical Considerations:
350
+ NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
351
+
352
+ Please report security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).
353
+
354
+ ### Example
355
+
356
+ The requirements are
357
+ ```
358
+ pip install qwen_vl_utils beir==2.0.0
359
+ ```
360
+
361
+ The dataset contains an [eval_example](https://huggingface.co/datasets/nvidia/miracl-vision/tree/main/eval_example) using [MrLight/dse-qwen2-2b-mrl-v1](https://huggingface.co/MrLight/dse-qwen2-2b-mrl-v1)​​
362
+
363
+ ```bash
364
+ python embedding_eval.py --dataset nvidia/miracl-vision --language en
365
+ ```
366
+
367
+ ### Loading the Dataset
368
+
369
+ ```python
370
+ from datasets import load_dataset
371
+
372
+ def hf_beir_queries(queries):
373
+ queries_beir = {}
374
+ for query in queries:
375
+ queries_beir[query['_id']] = query['text']
376
+ return(queries_beir)
377
+
378
+ def hf_beir_corpus(corpus):
379
+ corpus_beir = {}
380
+ for doc in corpus:
381
+ corpus_beir[doc['_id']] = doc
382
+ return(corpus_beir)
383
+
384
+ def hf_beir_qrels(qrels):
385
+ qrels_beir = {}
386
+ for el in qrels:
387
+ if str(el['query-id']) in qrels_beir:
388
+ qrels_beir[str(el['query-id'])][str(el['corpus-id'])] = el['score']
389
+ else:
390
+ qrels_beir[str(el['query-id'])] = {str(el['corpus-id']): el['score']}
391
+ return(qrels_beir)
392
+
393
+ def load_data(
394
+ path,
395
+ lang
396
+ ):
397
+ queries = load_dataset(path, 'queries-' + str(lang), split='default')
398
+ queries = hf_beir_queries(queries)
399
+ corpus = load_dataset(path, 'corpus-' + str(lang), split='default')
400
+ corpus = hf_beir_corpus(corpus)
401
+ qrels = load_dataset(path, 'qrels-' + str(lang), split='default')
402
+ qrels = hf_beir_qrels(qrels)
403
+ images = load_dataset(path, 'images-' + str(lang), split='default')
404
+ return(queries, corpus, qrels, images)
405
+
406
+ queries, corpus, qrels, images = load_data('nvidia/miracl-vision', 'en')
407
+ ```
408
+
409
+ ### Dataset Statistics
410
+
411
+ Number of Images: 338734
412
+
413
+ Number of questions: 7898
414
+
415
+ Total Data Storage: 95GB
416
+
417
+ | | | MIRACL (original) | MIRACL (original) | MIRACL-VISION | MIRACL-VISION |
418
+ |--------------|-------------------|:-----------------:|:------------------------:|:----------------:|:------------------:|
419
+ | **Language** | **Language Code** | **# of queries** | **# of document chunks** | **# of queries** | **# of documents** |
420
+ | Arabic | ar | 2896 | 2061414 | 2127 | 75444 |
421
+ | Bengali | bn | 411 | 297265 | 229 | 8495 |
422
+ | Chinese | zh | 393 | 4934368 | 189 | 8672 |
423
+ | English | en | 799 | 32893221 | 447 | 42971 |
424
+ | Farsi | fa | 632 | 2207172 | 342 | 15846 |
425
+ | Finnish | fi | 1271 | 1883509 | 791 | 33679 |
426
+ | French | fr | 343 | 14636953 | 142 | 6990 |
427
+ | German | de | 305 | 15866222 | 129 | 6302 |
428
+ | Hindi | hi | 350 | 506264 | 184 | 8004 |
429
+ | Indonesian | id | 960 | 1446315 | 603 | 23842 |
430
+ | Japanese | ja | 860 | 6953614 | 387 | 17909 |
431
+ | Korean | ko | 213 | 1486752 | 130 | 5700 |
432
+ | Russian | ru | 1252 | 9543918 | 564 | 25201 |
433
+ | Spanish | es | 648 | 10373953 | 369 | 17749 |
434
+ | Swahili | sw | 482 | 131924 | 239 | 7166 |
435
+ | Telugu | te | 828 | 518079 | 480 | 15429 |
436
+ | Thai | th | 733 | 542166 | 451 | 16313 |
437
+ | Yoruba | yo | 119 | 49043 | 95 | 3022 |
438
+ | | | | | | |
439
+ | **Avereage** | | **750** | **5907342** | **439** | **18819** |
440
+
441
+
442
+
443
+ ### Results
444
+
445
+ | | MIRACL-VISION (Text) | MIRACL-VISION (Text) | MIRACL-VISION (Text) | MIRACL-VISION (Text) | MIRACL-VISION (Image) | MIRACL-VISION (Image) | MIRACL-VISION (Image) | MIRACL-VISION (Image) |
446
+ |-------------------------|:-------------------------:|:---------------------------------:|:-------------------------:|:--------------------:|:-----------------------:|:----------------------------:|:---------------------:|:---------------------:|
447
+ | | **multilingual-e5-large** | **snowflake-arctic-embed-l-v2.0** | **gte-multilingual-base** | **bge-m3** | **dse-qwen2-2b-mrl-v1** | **gme-Qwen2-VL-2B-Instruct** | **vdr-2b-multi-v1** | **colqwen2-v1.0** |
448
+ | LLM Parameters (in M) | 560 | 567 | 305 | 567 | 1543 | 1543 | 1543 | 1543 |
449
+ | Language | | | | | | | | |
450
+ | Arabic | 0.8557 | 0.8754 | 0.8503 | 0.8883 | 0.3893 | 0.4888 | 0.4379 | 0.4129 |
451
+ | Bengali | 0.8421 | 0.8325 | 0.8211 | 0.8585 | 0.2352 | 0.3755 | 0.2473 | 0.2888 |
452
+ | Chinese | 0.6900 | 0.7179 | 0.7167 | 0.7458 | 0.5962 | 0.6314 | 0.5963 | 0.4926 |
453
+ | English | 0.7029 | 0.7437 | 0.7345 | 0.7348 | 0.6605 | 0.6784 | 0.6784 | 0.6417 |
454
+ | Farsi | 0.6793 | 0.7001 | 0.6984 | 0.7297 | 0.2250 | 0.3085 | 0.2398 | 0.2616 |
455
+ | Finnish | 0.8974 | 0.9014 | 0.8957 | 0.9071 | 0.4162 | 0.6863 | 0.5283 | 0.6604 |
456
+ | French | 0.7208 | 0.8236 | 0.7771 | 0.8158 | 0.7160 | 0.6851 | 0.7194 | 0.6876 |
457
+ | German | 0.7622 | 0.7774 | 0.7498 | 0.7695 | 0.6267 | 0.6345 | 0.6205 | 0.5995 |
458
+ | Hindi | 0.7595 | 0.7255 | 0.6916 | 0.7581 | 0.1740 | 0.3127 | 0.2058 | 0.2209 |
459
+ | Indonesian | 0.6793 | 0.6906 | 0.6757 | 0.7049 | 0.4866 | 0.5416 | 0.5254 | 0.5320 |
460
+ | Japanese | 0.8378 | 0.8484 | 0.8442 | 0.8720 | 0.6232 | 0.7305 | 0.6553 | 0.6970 |
461
+ | Korean | 0.7327 | 0.7545 | 0.7397 | 0.7934 | 0.4446 | 0.6202 | 0.4952 | 0.4419 |
462
+ | Russian | 0.7857 | 0.8242 | 0.8023 | 0.8363 | 0.6505 | 0.7202 | 0.6995 | 0.6811 |
463
+ | Spanish | 0.6596 | 0.7250 | 0.7029 | 0.7268 | 0.5927 | 0.6277 | 0.6274 | 0.6224 |
464
+ | Swahili | 0.8157 | 0.8089 | 0.7987 | 0.8337 | 0.4156 | 0.5348 | 0.4509 | 0.4931 |
465
+ | Telugu | 0.8948 | 0.9201 | 0.9076 | 0.9090 | 0.0274 | 0.0893 | 0.0318 | 0.0264 |
466
+ | Thai | 0.8424 | 0.8485 | 0.8509 | 0.8682 | 0.2692 | 0.3563 | 0.3177 | 0.2389 |
467
+ | Yoruba | 0.5655 | 0.5332 | 0.5698 | 0.5842 | 0.4178 | 0.4884 | 0.4577 | 0.5120 |
468
+ | | | | | | | | | |
469
+ | **Average** | **0.7624** | **0.7806** | **0.7682** | **0.7964** | **0.4426** | **0.5283** | **0.4741** | **0.4728** |
470
+ | **Average w/o Thelugu** | **0.7546** | **0.7724** | **0.7600** | **0.7898** | **0.4670** | **0.5542** | **0.5002** | **0.4991** |