File size: 27,397 Bytes
471c923
99f3b66
471c923
99f3b66
 
 
 
 
 
471c923
99f3b66
471c923
99f3b66
471c923
99f3b66
471c923
99f3b66
471c923
99f3b66
 
 
471c923
99f3b66
471c923
99f3b66
471c923
99f3b66
471c923
99f3b66
471c923
99f3b66
471c923
99f3b66
 
 
 
 
471c923
99f3b66
471c923
99f3b66
 
 
 
 
 
 
 
471c923
99f3b66
471c923
99f3b66
 
 
 
 
471c923
99f3b66
471c923
99f3b66
471c923
99f3b66
471c923
99f3b66
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
471c923
99f3b66
471c923
99f3b66
471c923
99f3b66
 
471c923
99f3b66
 
 
 
 
 
 
 
 
 
 
 
 
471c923
99f3b66
471c923
 
 
99f3b66
471c923
99f3b66
 
471c923
99f3b66
 
 
 
 
 
471c923
99f3b66
 
 
471c923
99f3b66
471c923
99f3b66
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
471c923
99f3b66
 
 
 
 
 
471c923
99f3b66
471c923
99f3b66
471c923
99f3b66
 
471c923
99f3b66
471c923
99f3b66
 
 
 
 
471c923
99f3b66
 
 
471c923
99f3b66
 
471c923
99f3b66
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
---
license: gemma
library_name: transformers
extra_gated_heading: Access PaliGemma on Hugging Face
extra_gated_prompt: To access PaliGemma on Hugging Face, you’re required to review
  and agree to Google’s usage license. To do this, please ensure you’re logged-in
  to Hugging Face and click below. Requests are processed immediately.
extra_gated_button_content: Acknowledge license
pipeline_tag: image-text-to-text
---
# PaliGemma model card

**Model page:** [PaliGemma](https://ai.google.dev/gemma/docs/paligemma)

Transformers PaliGemma 3B weights, pre-trained with 224*224 input images and 128 token input/output text sequences. The models are available in float32, bfloat16 and float16 formats for fine-tuning.

**Resources and technical documentation:**

*   [Responsible Generative AI Toolkit](https://ai.google.dev/responsible)
*   [PaliGemma on Kaggle](https://www.kaggle.com/models/google/paligemma)
*   [PaliGemma on Vertex Model Garden](https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/363)

**Terms of Use:** [Terms](https://ai.google.dev/gemma/terms)

**Authors:** Google

## Model information

### Model summary

#### Description

PaliGemma is a versatile and lightweight vision-language model (VLM) inspired by
[PaLI-3](https://arxiv.org/abs/2310.09199) and based on open components such as
the [SigLIP vision model](https://arxiv.org/abs/2303.15343) and the [Gemma
language model](https://arxiv.org/abs/2403.08295). It takes both image and text
as input and generates text as output, supporting multiple languages. It is designed for class-leading fine-tune performance on a wide range of vision-language tasks such as image and short video caption, visual question answering, text reading, object detection and object segmentation.

#### Model architecture

PaliGemma is the composition of a [Transformer
decoder](https://arxiv.org/abs/1706.03762) and a [Vision Transformer image
encoder](https://arxiv.org/abs/2010.11929), with a total of 3 billion
params. The text decoder is initialized from
[Gemma-2B](https://www.kaggle.com/models/google/gemma). The image encoder is
initialized from
[SigLIP-So400m/14](https://colab.research.google.com/github/google-research/big_vision/blob/main/big_vision/configs/proj/image_text/SigLIP_demo.ipynb).
PaliGemma is trained following the PaLI-3 recipes.

#### Inputs and outputs

*   **Input:** Image and text string, such as a prompt to caption the image, or
    a question.
*   **Output:** Generated text in response to the input, such as a caption of
    the image, an answer to a question, a list of object bounding box
    coordinates, or segmentation codewords.

### Model data

#### Pre-train datasets

PaliGemma is pre-trained on the following mixture of datasets:

*   **WebLI:** [WebLI (Web Language Image)](https://arxiv.org/abs/2209.06794) is
    a web-scale multilingual image-text dataset built from the public web. A
    wide range of WebLI splits are used to acquire versatile model capabilities,
    such as visual semantic understanding, object localization,
    visually-situated text understanding, multilinguality, etc.
*   **CC3M-35L:** Curated English image-alt_text pairs from webpages ([Sharma et
    al., 2018](https://aclanthology.org/P18-1238/)). We used the [Google Cloud
    Translation API](https://cloud.google.com/translate) to translate into 34
    additional languages.
*   **VQ²A-CC3M-35L/VQG-CC3M-35L:** A subset of VQ2A-CC3M ([Changpinyo et al.,
    2022a](https://aclanthology.org/2022.naacl-main.142/)), translated into the
    same additional 34 languages as CC3M-35L, using the [Google Cloud
    Translation API](https://cloud.google.com/translate).
*   **OpenImages:** Detection and object-aware questions and answers
    ([Piergiovanni et al. 2022](https://arxiv.org/abs/2209.04372)) generated by
    handcrafted rules on the [OpenImages dataset].
*   **WIT:** Images and texts collected from Wikipedia ([Srinivasan et al.,
    2021](https://arxiv.org/abs/2103.01913)).

[OpenImages dataset]: https://storage.googleapis.com/openimages/web/factsfigures_v7.html

#### Data responsibility filtering

The following filters are applied to WebLI, with the goal of training PaliGemma
on clean data:

*   **Pornographic image filtering:** This filter removes images deemed to be of
    pornographic nature.
*   **Text safety filtering:** We identify and filter out images that are paired
    with unsafe text. Unsafe text is any text deemed to contain or be about
    CSAI, pornography, vulgarities, or otherwise offensive.
*   **Text toxicity filtering:** We further use the [Perspective
    API](https://perspectiveapi.com/) to identify and filter out images that are
    paired with text deemed insulting, obscene, hateful or otherwise toxic.
*   **Text personal information filtering:** We filtered certain personal information and other sensitive data using [Cloud Data Loss Prevention (DLP)
    API](https://cloud.google.com/security/products/dlp) to protect the privacy
    of individuals. Identifiers such as social security numbers and [other sensitive information types] were removed.
*   **Additional methods:** Filtering based on content quality and safety in
    line with our policies and practices.

[other sensitive information types]: https://cloud.google.com/sensitive-data-protection/docs/high-sensitivity-infotypes-reference?_gl=1*jg604m*_ga*ODk5MzA3ODQyLjE3MTAzMzQ3NTk.*_ga_WH2QY8WWF5*MTcxMDUxNTkxMS4yLjEuMTcxMDUxNjA2NC4wLjAuMA..&_ga=2.172110058.-899307842.1710334759



## How to Use

PaliGemma is a single-turn vision language model not meant for conversational use,
and it works best when fine-tuning to a specific use case. 

You can configure which task the model will solve by conditioning it with task prefixes,
such as “detect” or “segment”. The pretrained models were trained in this fashion to imbue
them with a rich set of capabilities (question answering, captioning, segmentation, etc.).
However, they are not designed to be used directly, but to be transferred (by fine-tuning)
to specific tasks using a similar prompt structure. For interactive testing, you can use
the "mix" family of models, which have been fine-tuned on a mixture of tasks. 

Please, refer to the [usage and limitations section](#usage-and-limitations) for intended
use cases, or visit the [blog post](https://huggingface.co/blog/paligemma-google-vlm) for
additional details and examples.

## Use in Transformers

The following snippets use model `google/paligemma-3b-mix-224` for reference purposes.
The model in this repo you are now browsing may have been trained for other tasks, please
make sure you use appropriate inputs for the task at hand.

### Running the default precision (`float32`) on CPU

```python
from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
from PIL import Image
import requests
import torch

model_id = "google/paligemma-3b-mix-224"

url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg?download=true"
image = Image.open(requests.get(url, stream=True).raw)

model = PaliGemmaForConditionalGeneration.from_pretrained(model_id).eval()
processor = AutoProcessor.from_pretrained(model_id)

# Instruct the model to create a caption in Spanish
prompt = "caption es"
model_inputs = processor(text=prompt, images=image, return_tensors="pt")
input_len = model_inputs["input_ids"].shape[-1]

with torch.inference_mode():
    generation = model.generate(**model_inputs, max_new_tokens=100, do_sample=False)
    generation = generation[0][input_len:]
    decoded = processor.decode(generation, skip_special_tokens=True)
    print(decoded)
```

Output: `Un auto azul estacionado frente a un edificio.`

### Running other precisions on CUDA

For convenience, the repos contain revisions of the weights already converted to `bfloat16` and `float16`,
so you can use them to reduce the download size and avoid casting on your local computer.

This is how you'd run `bfloat16` on an nvidia CUDA card.

```python
from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
from PIL import Image
import requests
import torch

model_id = "google/paligemma-3b-mix-224"
device = "cuda:0"
dtype = torch.bfloat16

url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg?download=true"
image = Image.open(requests.get(url, stream=True).raw)

model = PaliGemmaForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=dtype,
    device_map=device,
    revision="bfloat16",
).eval()
processor = AutoProcessor.from_pretrained(model_id)

# Instruct the model to create a caption in Spanish
prompt = "caption es"
model_inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)
input_len = model_inputs["input_ids"].shape[-1]

with torch.inference_mode():
    generation = model.generate(**model_inputs, max_new_tokens=100, do_sample=False)
    generation = generation[0][input_len:]
    decoded = processor.decode(generation, skip_special_tokens=True)
    print(decoded)
```

### Loading in 4-bit / 8-bit

You need to install `bitsandbytes` to automatically run inference using 8-bit or 4-bit precision:

```
pip install bitsandbytes accelerate
```

```
from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
from PIL import Image
import requests
import torch

model_id = "google/paligemma-3b-mix-224"
device = "cuda:0"
dtype = torch.bfloat16

url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg?download=true"
image = Image.open(requests.get(url, stream=True).raw)

quantization_config = BitsAndBytesConfig(load_in_8bit=True)

model = PaliGemmaForConditionalGeneration.from_pretrained(
    model_id, quantization_config=quantization_config
).eval()
processor = AutoProcessor.from_pretrained(model_id)

# Instruct the model to create a caption in Spanish
prompt = "caption es"
model_inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)
input_len = model_inputs["input_ids"].shape[-1]

with torch.inference_mode():
    generation = model.generate(**model_inputs, max_new_tokens=100, do_sample=False)
    generation = generation[0][input_len:]
    decoded = processor.decode(generation, skip_special_tokens=True)
    print(decoded)
```

## Implementation information

### Hardware

PaliGemma was trained using the latest generation of Tensor Processing Unit
(TPU) hardware (TPUv5e).

### Software

Training was done using [JAX](https://github.com/google/jax),
[Flax](https://github.com/google/flax),
[TFDS](https://github.com/tensorflow/datasets) and
[`big_vision`](https://github.com/google-research/big_vision).

JAX allows researchers to take advantage of the latest generation of hardware,
including TPUs, for faster and more efficient training of large models.

TFDS is used to access datasets and Flax is used for model architecture. The
PaliGemma fine-tune code and inference code are released in the `big_vision`
GitHub repository.

## Evaluation information

### Benchmark results

In order to verify the transferability of PaliGemma to a wide variety of
academic tasks, we fine-tune the pretrained models on each task. Additionally we
train the mix model with a mixture of the transfer tasks. We report results on
different resolutions to provide an impression of which tasks benefit from
increased resolution. Importantly, none of these tasks or datasets are part of
the pretraining data mixture, and their images are explicitly removed from the
web-scale pre-training data.

#### Single task (fine-tune on single task)

<table>
  <tbody><tr>
    <th>Benchmark<br>(train split)</th>
    <th>Metric<br>(split)</th>
    <th>pt-224</th>
    <th>pt-448</th>
    <th>pt-896</th>
  </tr>
  <tr>
    <th>Captioning</th>
  </tr>
  <tr>
    <td>
      <a href="https://cocodataset.org/#home">COCO captions</a><br>(train+restval)
    </td>
    <td>CIDEr (val)</td>
    <td>141.92</td>
    <td>144.60</td>
  </tr>
  <tr>
    <td>
      <a href="https://nocaps.org/">NoCaps</a><br>(Eval of COCO<br>captions transfer)
    </td>
    <td>CIDEr (val)</td>
    <td>121.72</td>
    <td>123.58</td>
  </tr>
  <tr>
    <td>
      <a href="https://arxiv.org/pdf/2205.12522">COCO-35L</a><br>(train)
    </td>
    <td>CIDEr dev<br>(en/avg-34/avg)</td>
    <td>
      139.2<br>
      115.8<br>
      116.4
    </td>
    <td>
      141.2<br>
      118.0<br>
      118.6
    </td>
  </tr>
  <tr>
    <td>
      <a href="https://arxiv.org/pdf/2205.12522">XM3600</a><br>(Eval of COCO-35L transfer)
    </td>
    <td>CIDEr dev<br>(en/avg-34/avg)</td>
    <td>
      78.1<br>
      41.3<br>
      42.4
    </td>
    <td>
      80.0<br>
      41.9<br>
      42.9
    </td>
  </tr>
  <tr>
    <td>
      <a href="https://textvqa.org/textcaps/">TextCaps</a><br>(train)
    </td>
    <td>CIDEr (val)</td>
    <td>127.48</td>
    <td>153.94</td>
  </tr>
  <tr>
    <td>
      <a href="https://arxiv.org/abs/2110.11624">SciCap</a><br>(first sentence, no subfigure)<br>(train+val)
    </td>
    <td>CIDEr/BLEU-4<br>(test)</td>
    <td>
      162.25<br>
      0.192<br>
    </td>
    <td>
      181.49<br>
      0.211<br>
    </td>
  </tr>
  <tr>
    <td>
      <a href="https://arxiv.org/abs/2108.03353">Screen2words</a><br>(train+dev)
    </td>
    <td>CIDEr (test)</td>
    <td>117.57</td>
    <td>119.59</td>
  </tr>
  <tr>
    <td>
      <a href="https://arxiv.org/abs/2010.04295">Widget Captioning</a><br>(train+dev)
    </td>
    <td>CIDEr (test)</td>
    <td>136.07</td>
    <td>148.36</td>
  </tr>
  <tr>
    <th>Question answering</th>
  </tr>
  <tr>
    <td>
      <a href="https://visualqa.org/index.html">VQAv2</a><br>(train+validation)
    </td>
    <td>Accuracy<br>(Test server - std)</td>
    <td>83.19</td>
    <td>85.64</td>
  </tr>
  <tr>
    <td>
      <a href="https://arxiv.org/abs/2401.06209">MMVP</a><br>(Eval of VQAv2 transfer)
    </td>
    <td>Paired Accuracy</td>
    <td>47.33</td>
    <td>45.33</td>
  </tr>
  <tr>
    <td>
      <a href="https://arxiv.org/abs/2305.10355">POPE</a><br>(Eval of VQAv2 transfer)
    </td>
    <td>Accuracy<br>(random/popular/<br>adversarial)</td>
    <td>
      87.80<br>
      85.87<br>
      84.27
    </td>
    <td>
      88.23<br>
      86.77<br>
      85.90
    </td>
  </tr>
  <tr>
    <td>
      <a href="https://okvqa.allenai.org/">OKVQA</a><br>(train)
    </td>
    <td>Accuracy (val)</td>
    <td>63.54</td>
    <td>63.15</td>
  </tr>
  <tr>
    <td>
      <a href="https://allenai.org/project/a-okvqa/home">A-OKVQA</a> (MC)<br>(train+val)
    </td>
    <td>Accuracy<br>(Test server)</td>
    <td>76.37</td>
    <td>76.90</td>
  </tr>
  <tr>
    <td>
      <a href="https://allenai.org/project/a-okvqa/home">A-OKVQA</a> (DA)<br>(train+val)
    </td>
    <td>Accuracy<br>(Test server)</td>
    <td>61.85</td>
    <td>63.22</td>
  </tr>
  <tr>
    <td>
      <a href="https://cs.stanford.edu/people/dorarad/gqa/about.html">GQA</a><br>(train_balanced+<br>val_balanced)
    </td>
    <td>Accuracy<br>(testdev balanced)</td>
    <td>65.61</td>
    <td>67.03</td>
  </tr>
  <tr>
    <td>
      <a href="https://aclanthology.org/2022.findings-acl.196/">xGQA</a><br>(Eval of GQA transfer)
    </td>
    <td>Mean Accuracy<br>(bn, de, en, id,<br>ko, pt, ru, zh)</td>
    <td>58.37</td>
    <td>59.07</td>
  </tr>
  <tr>
    <td>
      <a href="https://lil.nlp.cornell.edu/nlvr/">NLVR2</a><br>(train+dev)
    </td>
    <td>Accuracy (test)</td>
    <td>90.02</td>
    <td>88.93</td>
  </tr>
  <tr>
    <td>
      <a href="https://marvl-challenge.github.io/">MaRVL</a><br>(Eval of NLVR2 transfer)
    </td>
    <td>Mean Accuracy<br>(test)<br>(id, sw, ta, tr, zh)</td>
    <td>80.57</td>
    <td>76.78</td>
  </tr>
  <tr>
    <td>
      <a href="https://allenai.org/data/diagrams">AI2D</a><br>(train)
    </td>
    <td>Accuracy (test)</td>
    <td>72.12</td>
    <td>73.28</td>
  </tr>
  <tr>
    <td>
      <a href="https://scienceqa.github.io/">ScienceQA</a><br>(Img subset, no CoT)<br>(train+val)
    </td>
    <td>Accuracy (test)</td>
    <td>95.39</td>
    <td>95.93</td>
  </tr>
  <tr>
    <td>
      <a href="https://zenodo.org/records/6344334">RSVQA-LR</a> (Non numeric)<br>(train+val)
    </td>
    <td>Mean Accuracy<br>(test)</td>
    <td>92.65</td>
    <td>93.11</td>
  </tr>
  <tr>
    <td>
      <a href="https://zenodo.org/records/6344367">RSVQA-HR</a> (Non numeric)<br>(train+val)
    </td>
    <td>Mean Accuracy<br>(test/test2)</td>
    <td>
      92.61<br>
      90.58
    </td>
    <td>
      92.79<br>
      90.54
    </td>
  </tr>
  <tr>
    <td>
      <a href="https://arxiv.org/abs/2203.10244">ChartQA</a><br>(human+aug)x(train+val)
    </td>
    <td>Mean Relaxed<br>Accuracy<br>(test_human,<br>test_aug)</td>
    <td>57.08</td>
    <td>71.36</td>
  </tr>
  <tr>
    <td>
      <a href="https://vizwiz.org/tasks-and-datasets/vqa/">VizWiz VQA</a><br>(train+val)
    </td>
    <td>Accuracy<br>(Test server - std)</td>
    <td>
      73.7
    </td>
    <td>
      75.52
    </td>
  </tr>
  <tr>
    <td>
      <a href="https://arxiv.org/abs/1810.12440">TallyQA</a><br>(train)
    </td>
    <td>Accuracy<br>(test_simple/<br>test_complex)</td>
    <td>
      81.72<br>
      69.56
    </td>
    <td>
      84.86<br>
      72.27
    </td>
  </tr>
  <tr>
    <td>
      <a href="https://ocr-vqa.github.io/">OCR-VQA</a><br>(train+val)
    </td>
    <td>Accuracy (test)</td>
    <td>72.32</td>
    <td>74.61</td>
    <td>74.93</td>
  </tr>
  <tr>
    <td>
      <a href="https://textvqa.org/">TextVQA</a><br>(train+val)
    </td>
    <td>Accuracy<br>(Test server - std)</td>
    <td>55.47</td>
    <td>73.15</td>
    <td>76.48</td>
  </tr>
  <tr>
    <td>
      <a href="https://www.docvqa.org/">DocVQA</a><br>(train+val)
    </td>
    <td>ANLS (Test server)</td>
    <td>43.74</td>
    <td>78.02</td>
    <td>84.77</td>
  </tr>
  <tr>
    <td>
      <a href="https://openaccess.thecvf.com/content/WACV2022/papers/Mathew_InfographicVQA_WACV_2022_paper.pdf">Infographic VQA</a><br>(train+val)
    </td>
    <td>ANLS (Test server)</td>
    <td>28.46</td>
    <td>40.47</td>
    <td>47.75</td>
  </tr>
  <tr>
    <td>
      <a href="https://arxiv.org/abs/1905.13648">SceneText VQA</a><br>(train+val)
    </td>
    <td>ANLS (Test server)</td>
    <td>63.29</td>
    <td>81.82</td>
    <td>84.40</td>
  </tr>
  <tr>
    <th>Segmentation</th>
  </tr>
  <tr>
    <td>
      <a href="https://arxiv.org/abs/1608.00272">RefCOCO</a><br>(combined refcoco, refcoco+,<br>refcocog excluding val<br>and test images)
    </td>
    <td>MIoU<br>(validation)<br>refcoco/refcoco+/<br>refcocog</td>
    <td>
      73.40<br>
      68.32<br>
      67.65
    </td>
    <td>
      75.57<br>
      69.76<br>
      70.17
    </td>
    <td>
      76.94<br>
      72.18<br>
      72.22
    </td>
  </tr>
  <tr>
    <th>Video tasks (Caption/QA)</th>
  </tr>
  <tr>
    <td>MSR-VTT (Captioning)</td>
    <td>CIDEr (test)</td>
    <td>70.54</td>
  </tr>
  <tr>
    <td>MSR-VTT (QA)</td>
    <td>Accuracy (test)</td>
    <td>50.09</td>
  </tr>
  <tr>
    <td>ActivityNet (Captioning)</td>
    <td>CIDEr (test)</td>
    <td>34.62</td>
  </tr>
  <tr>
    <td>ActivityNet (QA)</td>
    <td>Accuracy (test)</td>
    <td>50.78</td>
  </tr>
  <tr>
    <td>VATEX (Captioning)</td>
    <td>CIDEr (test)</td>
    <td>79.73</td>
  </tr>
  <tr>
    <td>MSVD (QA)</td>
    <td>Accuracy (test)</td>
    <td>60.22</td>
  </tr>
</tbody></table>

#### Mix model (fine-tune on mixture of transfer tasks)

<table>
  <tbody><tr>
    <th>Benchmark</th>
    <th>Metric (split)</th>
    <th>mix-224</th>
    <th>mix-448</th>
  </tr>
  <tr>
    <td><a href="https://arxiv.org/abs/2401.06209">MMVP</a></td>
    <td>Paired Accuracy</td>
    <td>46.00</td>
    <td>45.33</td>
  </tr>
  <tr>
    <td><a href="https://arxiv.org/abs/2305.10355">POPE</a></td>
    <td>Accuracy<br>(random/popular/adversarial)</td>
    <td>
      88.00<br>
      86.63<br>
      85.67
    </td>
    <td>
      89.37<br>
      88.40<br>
      87.47
    </td>
  </tr>
</tbody></table>

## Ethics and safety

### Evaluation approach

Our evaluation methods include structured evaluations and internal red-teaming
testing of relevant content policies. Red-teaming was conducted by a number of
different teams, each with different goals and human evaluation metrics. These
models were evaluated against a number of different categories relevant to
ethics and safety, including:

*   Human evaluation on prompts covering child safety, content safety and
    representational harms. See the [Gemma model
    card](https://ai.google.dev/gemma/docs/model_card#evaluation_approach) for
    more details on evaluation approach, but with image captioning and visual
    question answering setups.
*   Image-to-Text benchmark evaluation: Benchmark against relevant academic
    datasets such as FairFace Dataset ([Karkkainen et al.,
    2021](https://arxiv.org/abs/1908.04913)).

### Evaluation results

*   The human evaluation results of ethics and safety evaluations are within
    acceptable thresholds for meeting [internal
    policies](https://storage.googleapis.com/gweb-uniblog-publish-prod/documents/2023_Google_AI_Principles_Progress_Update.pdf#page=11)
    for categories such as child safety, content safety and representational
    harms.
*   On top of robust internal evaluations, we also use the Perspective API
    (threshold of 0.8) to measure toxicity, profanity, and other potential
    issues in the generated captions for images sourced from the FairFace
    dataset. We report the maximum and median values observed across subgroups
    for each of the perceived gender, ethnicity, and age attributes.


<table>
  <tbody><tr>
    </tr></tbody><tbody><tr><th>Metric</th>
    <th>Perceived<br>gender</th>
    <th></th>
    <th>Ethnicity</th>
    <th></th>
    <th>Age group</th>
    <th></th>
  </tr>
  <tr>
    <th></th>
    <th>Maximum</th>
    <th>Median</th>
    <th>Maximum</th>
    <th>Median</th>
    <th>Maximum</th>
    <th>Median</th>
  </tr>
  <tr>
    <td>Toxicity</td>
    <td>0.04%</td>
    <td>0.03%</td>
    <td>0.08%</td>
    <td>0.00%</td>
    <td>0.09%</td>
    <td>0.00%</td>
  </tr>
  <tr>
    <td>Identity Attack</td>
    <td>0.00%</td>
    <td>0.00%</td>
    <td>0.00%</td>
    <td>0.00%</td>
    <td>0.00%</td>
    <td>0.00%</td>
  </tr>
  <tr>
    <td>Insult</td>
    <td>0.06%</td>
    <td>0.04%</td>
    <td>0.09%</td>
    <td>0.07%</td>
    <td>0.16%</td>
    <td>0.00%</td>
  </tr>
  <tr>
    <td>Threat</td>
    <td>0.06%</td>
    <td>0.05%</td>
    <td>0.14%</td>
    <td>0.05%</td>
    <td>0.17%</td>
    <td>0.00%</td>
  </tr>
  <tr>
    <td>Profanity</td>
    <td>0.00%</td>
    <td>0.00%</td>
    <td>0.00%</td>
    <td>0.00%</td>
    <td>0.00%</td>
    <td>0.00%</td>
  </tr>
</tbody></table>

## Usage and limitations

### Intended usage

Open Vision Language Models (VLMs) have a wide range of applications across
various industries and domains. The following list of potential uses is not
comprehensive. The purpose of this list is to provide contextual information
about the possible use-cases that the model creators considered as part of model
training and development.

Fine-tune on specific vision-language task:

*   The pre-trained models can be fine-tuned on a wide range of vision-language
    tasks such as: image captioning, short video caption, visual question
    answering, text reading, object detection and object segmentation.
*   The pre-trained models can be fine-tuned for specific domains such as remote
    sensing question answering, visual questions from people who are blind,
    science question answering, describe UI element functionalities.
*   The pre-trained models can be fine-tuned for tasks with non-textual outputs
    such as bounding boxes or segmentation masks.

Vision-language research:

*   The pre-trained models and fine-tuned models can serve as a foundation for researchers to experiment with VLM
    techniques, develop algorithms, and contribute to the advancement of the
    field.

### Ethical considerations and risks

The development of vision-language models (VLMs) raises several ethical concerns. In creating an open model, we have carefully considered the following:

*   Bias and Fairness
    *   VLMs trained on large-scale, real-world image-text data can reflect socio-cultural biases embedded in the training material. These models underwent careful scrutiny, input data pre-processing described and posterior evaluations reported in this card.
*   Misinformation and Misuse
    *   VLMs can be misused to generate text that is false, misleading, or harmful.
    *   Guidelines are provided for responsible use with the model, see the [Responsible Generative AI Toolkit](https://ai.google.dev/responsible).
*   Transparency and Accountability
    *   This model card summarizes details on the models' architecture, capabilities, limitations, and evaluation processes.
    *   A responsibly developed open model offers the opportunity to share innovation by making VLM technology accessible to developers and researchers across the AI ecosystem.


Risks identified and mitigations:

*   **Perpetuation of biases:** It's encouraged to perform continuous monitoring
    (using evaluation metrics, human review) and the exploration of de-biasing
    techniques during model training, fine-tuning, and other use cases.
*   **Generation of harmful content:** Mechanisms and guidelines for content
    safety are essential. Developers are encouraged to exercise caution and
    implement appropriate content safety safeguards based on their specific
    product policies and application use cases.
*   **Misuse for malicious purposes:** Technical limitations and developer and
    end-user education can help mitigate against malicious applications of LLMs.
    Educational resources and reporting mechanisms for users to flag misuse are
    provided. Prohibited uses of Gemma models are outlined in the [Gemma
    Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy).
*   **Privacy violations:** Models were trained on data filtered to remove certain personal information and sensitive data. Developers are encouraged to adhere to privacy regulations with privacy-preserving techniques.

### Limitations

*   Most limitations inherited from the underlying Gemma model still apply:
    *   VLMs are better at tasks that can be framed with clear prompts and
        instructions. Open-ended or highly complex tasks might be challenging.
    *   Natural language is inherently complex. VLMs might struggle to grasp
        subtle nuances, sarcasm, or figurative language.
    *   VLMs generate responses based on information they learned from their
        training datasets, but they are not knowledge bases. They may generate
        incorrect or outdated factual statements.
    *   VLMs rely on statistical patterns in language and images. They might
        lack the ability to apply common sense reasoning in certain situations.
*   PaliGemma was designed first and foremost to serve as a general pre-trained
    model for transfer to specialized tasks. Hence, its "out of the box" or
    "zero-shot" performance might lag behind models designed specifically for
    that.
*   PaliGemma is not a multi-turn chatbot. It is designed for a single round of
    image and text input.