Newspaper OCR Gold: 13,371 Verified Line Transcriptions from 1840s-2010s

13,371 line-level transcriptions verified by Qwen3-VL 235B form gold-standard OCR training data. The dataset includes line crop PNG images from 100 newspaper pages across 73 unique titles spanning the 1840s to the 2010s. It was created by NealCaren and is split by page into train, validation, and test sets.

Use Cases

• Train optical character recognition models based on the verified line-level transcriptions.
• Benchmark OCR accuracy on historical newspaper scans based on the gold-standard test split.
• Fine-tune vision-language models for document understanding based on the paired image-text data.
• Analyze OCR error patterns across different time periods based on the 1840s-2010s date range.

Strengths

• 13,371 verified line-level transcriptions provide a substantial training corpus.
• 100 newspaper pages with a train/val/test split (80/10/10) facilitates model development.
• 73 unique titles spanning the 1840s-2010s offer temporal diversity.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset's geographic and linguistic scope is not specified, which may indicate source bias.
• Row count for the main metadata file is unknown, which may limit suitability assessment.

Provenance

Source

huggingface

Collection Method

Line crop PNG images from newspaper scans with transcriptions verified by Qwen3-VL 235B.

Time Range

1840s-2010s

Freshness

Last updated 2026-03-23 14:00:02; freshness should be verified.