Dataset and Graph Resources
This repository contains the generated training and test datasets, knowledge graphs, and reference files.
π File Descriptions
Dataset Splits (Parquet)
The following files contain the generated dataset splits:
train.parquet: The training set generated for this project.seen_test.parquet: A test set created without specific filtering.unseen_test.parquet: A test set where documents do not overlap with the training set.
Since we use in-batch negatives, we simply removed all training samples containing documents that overlap with the unseen questions.
Knowledge Graph
one_hop_neighborhood_graph.json: This file contains the entity graph relations (1-hop neighborhood) specifically used for our dataset generation process.
Image Metadata (MuKA)
The following files contain web image URLs provided by MuKA:
evqa_passages_image_urls.jsonlinfoseek_passages_image_urls.jsonl
Images
images.zip: Contains the compressed image files used in this dataset.
We need to add paths and google lens searched images
π Data Sources & Acknowledgments
M2KR
The document passages used in this dataset are based on the M2KR dataset.
Important Note regarding Images: The images for datasets such as EVQA, OVEN and InfoSeek are contributions of the original M2KR paper.
- We rely on their contribution for these resources.
- Please visit the M2KR Project Page to access the original data and for further details.
- Downloads last month
- 7