metadata

library_name: transformers
license: gemma
base_model: vidore/colpaligemma-3b-pt-448-base
tags:
  - colpali
  - generated_from_trainer
model-index:
  - name: finetune_colpali_v1_2-ufo-4bit
    results: []
datasets:
  - davanstrien/ufo-ColPali

finetune_colpali_v1_2-ufo-4bit

This model is a fine-tuned version of vidore/colpaligemma-3b-pt-448-base on the davanstrien/ufo-ColPali dataset.

The model was trained using the fine tuning notebook from tonywu71. I changed almost nothing except the data processing steps.

The dataset used for training was created using synthetic data from Qwen/Qwen2-VL-7B-Instruct. The process for making this dataset is discussed more in the blog post.

The model achieves the following results on the evaluation set:

Loss: 0.1064
Model Preparation Time: 0.0056

Model description

This model is a fine tune of a ColPali vidore/colpaligemma-3b-pt-448-base:

ColPali 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 PaliGemma-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.

Intended uses & limitations

For retrieving UFO newsletters documents.

Training and evaluation data

The training data was created via the following steps:

Downloading a sample of UFO newsletters from this Internet archive Collection.
Using the pdf-to-page-images-dataset Space to convert the PDF documents into a single page image dataset
Use a VLM to generate synthetic queries for these documents using the approach outlines here. This results in davanstrien/ufo-ColPali.
Train the model using the fine tuning notebook from tonywu71. I changed almost nothing except the data processing steps.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 5e-05
train_batch_size: 4
eval_batch_size: 4
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 8
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 100
num_epochs: 1.5

Training results

Training Loss	Epoch	Step	Validation Loss	Model Preparation Time
No log	0.0041	1	0.1879	0.0056
0.1193	0.4090	100	0.1136	0.0056
0.1287	0.8180	200	0.1122	0.0056
0.0662	1.2270	300	0.1063	0.0056

Framework versions

Transformers 4.44.2
Pytorch 2.4.1+cu121
Datasets 3.0.0
Tokenizers 0.19.1