Colpali V1.3

🚀 ColPali: Visual Retriever based on PaliGemma-3B with ColBERT strategy

ColPali is a model that leverages a novel architecture and training strategy based on Vision Language Models (VLMs). It can efficiently index documents using their visual features. It extends PaliGemma - 3B to generate ColBERT-style multi - vector representations of text and images. This model was introduced in the paper ColPali: Efficient Document Retrieval with Vision Language Models and first released in this repository.

🚀 Quick Start

This version is trained with 256 batch size for 3 epochs on the same data as the original ColPali model.

✨ Features

Version specificity

This version is trained with colpali - engine==0.2.0 but can be loaded for any version >=0.2.0. Compared to vidore/colpali, it uses right padding for queries to fix unwanted tokens in the query encoding. It also stems from the fixed vidore/colpaligemma - 3b - pt - 448 - base to ensure deterministic projection layer initialization. It was trained for 5 epochs, with in - batch negatives and hard mined negatives and a warmup of 1000 steps (10x longer) to reduce non - English language collapse. The data used is the same as the ColPali data described in the paper.

Model Description

This model is built iteratively starting from an off - the - shelf [SigLIP](https://huggingface.co/google/siglip - so400m - patch14 - 384) model. We finetuned it to create BiSigLIP and fed the patch - embeddings output by SigLIP to an LLM, [PaliGemma - 3B](https://huggingface.co/google/paligemma - 3b - mix - 448) to create BiPali. Inputting image patch embeddings through a language model maps them to a latent space similar to textual input (query), enabling the use of the ColBERT strategy to compute interactions between text tokens and image patches, which significantly improves performance compared to BiPali.

📦 Installation

Install [colpali - engine](https://github.com/illuin - tech/colpali):

pip install colpali - engine>=0.3.0,<0.4.0

💻 Usage Examples

Basic Usage

Then run the following code:

from typing import cast

import torch
from PIL import Image

from colpali_engine.models import ColPali, ColPaliProcessor

model_name = "vidore/colpali-v1.3"

model = ColPali.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="cuda:0",  # or "mps" if on Apple Silicon
).eval()

processor = ColPaliProcessor.from_pretrained(model_name)

# Your inputs
images = [
    Image.new("RGB", (32, 32), color="white"),
    Image.new("RGB", (16, 16), color="black"),
]
queries = [
    "Is attention really all you need?",
    "Are Benjamin, Antoine, Merve, and Jo best friends?",
]

# Process the inputs
batch_images = processor.process_images(images).to(model.device)
batch_queries = processor.process_queries(queries).to(model.device)

# Forward pass
with torch.no_grad():
    image_embeddings = model(**batch_images)
    query_embeddings = model(**batch_queries)

scores = processor.score_multi_vector(query_embeddings, image_embeddings)

📚 Documentation

Model Training

Dataset

Our training dataset consists of 127,460 query - page pairs. 63% of it comes from openly available academic datasets, and 37% is a synthetic dataset composed of pages from web - crawled PDF documents with pseudo - questions generated by Claude - 3 Sonnet. The training set is fully English, allowing us to study zero - shot generalization to non - English languages. We ensure no multi - page PDF document is used in both [ViDoRe](https://huggingface.co/collections/vidore/vidore - benchmark - 667173f98e70a1c0fa4db00d) and the train set to avoid evaluation contamination. A validation set is created with 2% of the samples for hyperparameter tuning.

Note: Multilingual data is present in the pretraining corpus of the language model (Gemma - 2B) and may occur during PaliGemma - 3B's multimodal training.

Parameters

All models are trained for 1 epoch on the train set. Unless otherwise specified, we train models in bfloat16 format, use low - rank adapters (LoRA) with alpha = 32 and r = 32 on the transformer layers of the language model and the final randomly initialized projection layer, and use a paged_adamw_8bit optimizer. We train on an 8 - GPU setup with data parallelism, a learning rate of 5e - 5 with linear decay and 2.5% warmup steps, and a batch size of 32.

Limitations

• Focus: The model mainly focuses on PDF - type documents and high - resource languages, which may limit its generalization to other document types or less represented languages.
• Support: The model relies on multi - vector retrieval from the ColBERT late interaction mechanism, which may require engineering efforts to adapt to widely used vector retrieval frameworks lacking native multi - vector support.

License

ColPali's vision language backbone model (PaliGemma) is under the gemma license as specified in its [model card](https://huggingface.co/google/paligemma - 3b - mix - 448). The adapters attached to the model are under the MIT license.

Contact

• Manuel Faysse: manuel.faysse@illuin.tech
• Hugues Sibille: hugues.sibille@illuin.tech
• Tony Wu: tony.wu@illuin.tech

Citation

If you use any datasets or models from this organization in your research, please cite the original dataset as follows:

@misc{faysse2024colpaliefficientdocumentretrieval,
  title={ColPali: Efficient Document Retrieval with Vision Language Models}, 
  author={Manuel Faysse and Hugues Sibille and Tony Wu and Bilel Omrani and Gautier Viaud and Céline Hudelot and Pierre Colombo},
  year={2024},
  eprint={2407.01449},
  archivePrefix={arXiv},
  primaryClass={cs.IR},
  url={https://arxiv.org/abs/2407.01449}, 
}

📄 Information Table