PaliGemma model card

Model page: PaliGemma

JAX/FLAX PaliGemma 3B weights, fine-tuned with 224*224 input images on the GQA dataset. The models are available in float32, bfloat16 and float16 format for research purposes only. The fine-tune config is available at big_vision.

Resources and technical documentation:

Terms of Use: Terms

Authors: Google

Model information

Model summary

Description

PaliGemma is a versatile and lightweight vision-language model (VLM) inspired by PaLI-3 and based on open components such as the SigLIP vision model and the Gemma language model. 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 and a Vision Transformer image encoder, with a total of 3 billion params. The text decoder is initialized from Gemma-2B. The image encoder is initialized from SigLIP-So400m/14. 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) 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). We used the Google Cloud Translation API to translate into 34 additional languages.
VQ²A-CC3M-35L/VQG-CC3M-35L: A subset of VQ2A-CC3M (Changpinyo et al., 2022a), translated into the same additional 34 languages as CC3M-35L, using the Google Cloud Translation API.
OpenImages: Detection and object-aware questions and answers (Piergiovanni et al. 2022) generated by handcrafted rules on the OpenImages dataset.
WIT: Images and texts collected from Wikipedia (Srinivasan et al., 2021).

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 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 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.

Implementation information

Hardware

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

Software

Training was done using JAX, Flax, TFDS and 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.

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

Benchmark	Metric (split)	mix-224	mix-448
MMVP	Paired Accuracy	46.00	45.33
POPE	Accuracy (random/popular/adversarial)	88.00 86.63 85.67	89.37 88.40 87.47
GQA	Accuracy (test)	65.20	65.47

Single task (fine-tune on single task)

Benchmark (train split)	Metric (split)	pt-224	pt-448	pt-896
Captioning
COCO captions (train+restval)	CIDEr (val)	141.92	144.60
NoCaps (Eval of COCO captions transfer)	CIDEr (val)	121.72	123.58
COCO-35L (train)	CIDEr dev (en/avg-34/avg)	139.2 115.8 116.4	141.2 118.0 118.6
XM3600 (Eval of COCO-35L transfer)	CIDEr dev (en/avg-34/avg)	78.1 41.3 42.4	80.0 41.9 42.9
TextCaps (train)	CIDEr (val)	127.48	153.94
SciCap (first sentence, no subfigure) (train+val)	CIDEr/BLEU-4 (test)	162.25 0.192	181.49 0.211
Screen2words (train+dev)	CIDEr (test)	117.57	119.59
Widget Captioning (train+dev)	CIDEr (test)	136.07	148.36
Question answering
VQAv2 (train+validation)	Accuracy (Test server - std)	83.19	85.64
MMVP (Eval of VQAv2 transfer)	Paired Accuracy	47.33	45.33
POPE (Eval of VQAv2 transfer)	Accuracy (random/popular/ adversarial)	87.80 85.87 84.27	88.23 86.77 85.90
OKVQA (train)	Accuracy (val)	63.54	63.15
A-OKVQA (MC) (train+val)	Accuracy (Test server)	76.37	76.90
A-OKVQA (DA) (train+val)	Accuracy (Test server)	61.85	63.22
GQA (train_balanced+ val_balanced)	Accuracy (testdev balanced)	65.61	67.03
xGQA (Eval of GQA transfer)	Mean Accuracy (bn, de, en, id, ko, pt, ru, zh)	58.37	59.07
NLVR2 (train+dev)	Accuracy (test)	90.02	88.93
MaRVL (Eval of NLVR2 transfer)	Mean Accuracy (test) (id, sw, ta, tr, zh)	80.57	76.78
AI2D (train)	Accuracy (test)	72.12	73.28
ScienceQA (Img subset, no CoT) (train+val)	Accuracy (test)	95.39	95.93
RSVQA-LR (Non numeric) (train+val)	Mean Accuracy (test)	92.65	93.11
RSVQA-HR (Non numeric) (train+val)	Mean Accuracy (test/test2)	92.61 90.58	92.79 90.54
ChartQA (human+aug)x(train+val)	Mean Relaxed Accuracy (test_human, test_aug)	57.08	71.36
VizWiz VQA (train+val)	Accuracy (Test server - std)	73.7	75.52
TallyQA (train)	Accuracy (test_simple/ test_complex)	81.72 69.56	84.86 72.27
OCR-VQA (train+val)	Accuracy (test)	72.32	74.61	74.93
TextVQA (train+val)	Accuracy (Test server - std)	55.47	73.15	76.48
DocVQA (train+val)	ANLS (Test server)	43.74	78.02	84.77
Infographic VQA (train+val)	ANLS (Test server)	28.46	40.47	47.75
SceneText VQA (train+val)	ANLS (Test server)	63.29	81.82	84.40
Segmentation
RefCOCO (combined refcoco, refcoco+, refcocog excluding val and test images)	MIoU (validation) refcoco/refcoco+/ refcocog	73.40 68.32 67.65	75.57 69.76 70.17	76.94 72.18 72.22
Video tasks (Caption/QA)
MSR-VTT (Captioning)	CIDEr (test)	70.54
MSR-VTT (QA)	Accuracy (test)	50.09
ActivityNet (Captioning)	CIDEr (test)	34.62
ActivityNet (QA)	Accuracy (test)	50.78
VATEX (Captioning)	CIDEr (test)	79.73
MSVD (QA)	Accuracy (test)	60.22

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 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).

Evaluation results

The human evaluation results of ethics and safety evaluations are within acceptable thresholds for meeting internal policies 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.

Metric	Perceived gender		Ethnicity		Age group
	Maximum	Median	Maximum	Median	Maximum	Median
Toxicity	0.04%	0.03%	0.08%	0.00%	0.09%	0.00%
Identity Attack	0.00%	0.00%	0.00%	0.00%	0.00%	0.00%
Insult	0.06%	0.04%	0.09%	0.07%	0.16%	0.00%
Threat	0.06%	0.05%	0.14%	0.05%	0.17%	0.00%
Profanity	0.00%	0.00%	0.00%	0.00%	0.00%	0.00%

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.
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.
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.

Citation

@article{beyer2024paligemma,
    title={{PaliGemma: A versatile 3B VLM for transfer}},
    author={Lucas Beyer* and Andreas Steiner* and André Susano Pinto* and Alexander Kolesnikov* and Xiao Wang* and Daniel Salz and Maxim Neumann and Ibrahim Alabdulmohsin and Michael Tschannen and Emanuele Bugliarello and Thomas Unterthiner and Daniel Keysers and Skanda Koppula and Fangyu Liu and Adam Grycner and Alexey Gritsenko and Neil Houlsby and Manoj Kumar and Keran Rong and Julian Eisenschlos and Rishabh Kabra and Matthias Bauer and Matko Bošnjak and Xi Chen and Matthias Minderer and Paul Voigtlaender and Ioana Bica and Ivana Balazevic and Joan Puigcerver and Pinelopi Papalampidi and Olivier Henaff and Xi Xiong and Radu Soricut and Jeremiah Harmsen and Xiaohua Zhai*},
    year={2024},
    journal={arXiv preprint arXiv:2407.07726}
}

Find the paper here.

google
/

paligemma-3b-ft-gqa-224-jax

Access PaliGemma on Hugging Face