ColQwen2.5-3b-multilingual-v1.0: Multilingual Visual Retriever based on Qwen2.5-VL-3B-Instruct with ColBERT strategy

Name: colqwen2.5-3b-multilingual-v1.0-merged
Availability: InStock

This is the base version trained on 8xH100 80GB with per_device_batch_size=128 for 8 epoch.

ColQwen is a model based on a novel model architecture and training strategy based on Vision Language Models (VLMs) to efficiently index documents from their visual features. It is a Qwen2.5-VL-3B extension that generates ColBERT- style multi-vector representations of text and images. It was introduced in the paper ColPali: Efficient Document Retrieval with Vision Language Models and first released in this repository

Version specificity

This model takes dynamic image resolutions in input and does not resize them, changing their aspect ratio as in ColPali. Maximal resolution is set so that 768 image patches are created at most. Experiments show clear improvements with larger amounts of image patches, at the cost of memory requirements.

This version is trained with colpali-engine==0.3.9.

Data

German & English: Taken from the tsystems/vqa_de_en_batch1 dataset.
Multilingual dataset: Taken from llamaindex/vdr-multilingual-train.
Synthetic data: Taken from openbmb/VisRAG-Ret-Train-Synthetic-data dataset.
In-domain VQA dataset: Taken from openbmb/VisRAG-Ret-Train-In-domain-data dataset.
Colpali dataset: Taken from vidore/colpali_train_set.

Model Training

Parameters

We train models use low-rank adapters (LoRA) with alpha=128 and r=128 on the transformer layers from the language model, as well as the final randomly initialized projection layer, and use a paged_adamw_8bit optimizer. We train on an 8xH100 GPU setup with distributed data parallelism (via accelerate), a learning rate of 2e-4 with linear decay with 1% warmup steps, batch size per device is 128 in bfloat16 format

Installation

bash

pip install git+https://github.com/illuin-tech/colpali
pip install transformers==4.49.0
pip install flash-attn --no-build-isolation

Usage

Using Sentence Transformers

The model also loads directly as a Sentence Transformers MultiVectorEncoder, which exposes the familiar encode_query / encode_document / similarity API and computes the late interaction MaxSim scores for you:

pip install "sentence-transformers[image]>=6.0.0"

python

from sentence_transformers import MultiVectorEncoder

model = MultiVectorEncoder("tsystems/colqwen2.5-3b-multilingual-v1.0-merged")

queries = [
    "What is the variable represented on the y-axis of the graph?",
    "Total outlay is maximum in which year?",
]
documents = [
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc1.jpg",
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc2.jpg",
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc3.jpg",
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc4.jpg",
]

query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings[0].shape, document_embeddings[0].shape)
# torch.Size([25, 128]) torch.Size([755, 128])

scores = model.similarity(query_embeddings, document_embeddings)
print(scores)
# tensor([[14.0508,  8.3906,  9.0303,  8.1934],
#         [ 6.7480, 14.2441,  6.0889,  4.7104]], device='cuda:0')
print("Best document per query:", scores.argmax(dim=1))
# Best document per query: tensor([0, 1], device='cuda:0')

Documents may be passed as URLs, local paths or PIL.Image objects. Text given to encode_document is embedded as a plain passage, without the query augmentation tokens. The scores above come from the default bfloat16 load.

Using ColPali Engine

[!WARNING] Note: current colpali-engine no longer sends the Query: prefix that this checkpoint was trained with. ColQwen2_5_Processor carried it through 0.3.12 and it was dropped in 0.3.13 (illuin-tech/colpali#339). The Sentence Transformers configuration in this repository reproduces the original training-time format, so its embeddings differ slightly from current colpali-engine output. To restore the training format on the colpali-engine path, set processor.query_prefix = "Query: " before calling process_queries.

python

import torch
from PIL import Image

from colpali_engine.models import ColQwen2_5, ColQwen2_5_Processor

model = ColQwen2_5.from_pretrained(
        "tsystems/colqwen2.5-3b-multilingual-v1.0",
        torch_dtype=torch.bfloat16,
        device_map="cuda:0",  # or "mps" if on Apple Silicon
    ).eval()
processor = ColQwen2_5_Processor.from_pretrained("tsystems/colqwen2.5-3b-multilingual-v1.0")

# Your inputs
images = [
    Image.new("RGB", (32, 32), color="white"),
    Image.new("RGB", (16, 16), color="black"),
]
queries = [
    "Is attention really all you need?",
    "What is the amount of bananas farmed in Salvador?",
]

# 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)

Limitations

Focus: The model primarily focuses on PDF-type documents and high-ressources languages, potentially limiting its generalization to other document types or less represented languages.
Support: The model relies on multi-vector retreiving derived from the ColBERT late interaction mechanism, which may require engineering efforts to adapt to widely used vector retrieval frameworks that lack native multi-vector support.

License

ColQwen2.5's vision language backbone model (Qwen2.5-VL) is under apache2.0 license. The adapters attached to the model are under MIT license.

Citation

If you use this models from this organization in your research, please cite the original paper as follows:

bibtex

@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}, 
}

Developed by: T-Systems International