ce-amtic's picture
Update README.md
54591c4 verified
metadata
license: cc-by-4.0

ProcVQA-20M Annotations

Project Page | arXiv | Code | Model | Media

This repository contains the text annotations for the ProcVQA-20M dataset. The full image files are hosted separately on ProcVQA-20M-media.

Overview

This dataset is constructed from over 26 embodied datasets, comprising:

  • 20M QA pairs for training
  • 330K original trajectories
  • 50M annotated frames from ~5,000 hours of manipulation data
  • 200+ different tasks

Dataset Structure

The dataset is organized as follows:

.
├── images                  # Image file (only for the test and validation sets)
└── jsons
    ├── stage1              # Large-scale QA pairs without human selection, used in the first stage of training
    ├── stage2              # QA pairs generated with human selection, used in the second stage of training
    ├── test                # ID test set
    └── valid               # OOD validation set

To download the full set of image files, please refer to ProcVQA-20M-media on Modelscope.

Statistics

Name Mean Tokens Max Tokens Tokens Counts
stage1/trun8192 1226 8191 21.21 B 17.30 M
stage1/ext16384 2637 16384 1.60 B 605.1 K
stage2/trun8192 845 8191 2.76 B 3.27 M
stage2/ext16384 1574 16383 166.44 M 105.7 K
test/trun8192 838 8166 15.89 M 19.0 K
test/full 846 15751 16.07 M 19.0 K
valid/trun8192 1992 8165 34.86 M 17.5 K
valid/full 2034 16064 35.76 M 17.6 K

Sub-task Statistics (Stage 2)

total: 13688, avg: 2.8732466393921685, max: 24, min: 1, median: 3
distribution:
1: 1455     2: 5303     3: 4187     4: 1523     5: 210      6: 479
7: 123      8: 197      9: 19       10: 153     11: 2       12: 15
14: 8       15: 1       16: 9       18: 2       20: 1       24: 1