newspaper-ocr-gold

Gold-standard OCR training data for historical newspaper scans.

Contents

• 13,371 line-level transcriptions verified by Qwen3-VL 235B
• Line crop PNG images from 100 newspaper pages
• 73 unique titles spanning 1840s-2010s
• Train/val/test split by page (80/10/10)

Splits

Split	Pages	Lines
train	80	11,044
val	10	1,111
test	10	1,216

Files

• verified_lines.jsonl — full metadata (split, page_id, line_id, crop_path, transcription, confidence, flag)
• sample_metadata.json — page sampling details (73 titles, decade distribution)
• train_images.tar.gz, val_images.tar.gz, test_images.tar.gz — line crop PNGs organized as {split}/{page_id}/lines/line_NNNN.png

Usage

from huggingface_hub import hf_hub_download
import tarfile, json

# Download verified labels
path = hf_hub_download("NealCaren/newspaper-ocr-gold", "verified_lines.jsonl", repo_type="dataset")
with open(path) as f:
    lines = [json.loads(l) for l in f]

# Download and extract images for a split
for split in ["train", "val", "test"]:
    tar = hf_hub_download("NealCaren/newspaper-ocr-gold", f"{split}_images.tar.gz", repo_type="dataset")
    with tarfile.open(tar) as t:
        t.extractall("./gold_data/")

Quality

• 49% clean, 47% partial (line cut-off at word boundary), 3% degraded
• Mean confidence: 0.95
• Verified by Qwen3-VL 235B via OpenRouter (blind transcription, no OCR input)