UniDoc-Bench: 1,700 Multimodal QA Pairs from 70,000 PDF Pages

Salesforce developed UniDoc-Bench in 2024 as a benchmark for multimodal retrieval-augmented generation (MM-RAG). It contains 1,700+ multimodal QA pairs derived from a corpus of 70,000 real-world PDF pages across eight domains. The data links evidence across text, tables, and figures to support complex document-based reasoning tasks.

Use Cases

• Evaluating MM-RAG systems using the 1,700+ multimodal QA pairs to retrieve evidence from PDF pages
• Benchmarking Visual Question Answering (VQA) by interpreting data within the 'tables' and 'figures' features
• Training retrieval models to identify relevant 'text' segments within a 70,000-page document corpus

Strengths

• 70,000 real-world PDF pages providing a large-scale retrieval corpus
• 1,700+ multimodal QA pairs
• Covers eight distinct domains for cross-domain evaluation
• Links evidence across text, tables, and figures

Limitations

• Relatively small QA pair count (1,700+) compared to the 70,000-page document corpus size
• Limited to English language content
• Non-commercial license restricts industrial application

Provenance

Source

Salesforce

Collection Method

Extracted and linked from 70,000 real-world PDF pages across eight domains

Freshness

Last updated December 2025

Geography

United States