wjbmattingly/nara-qwen-3.5-2b

VISIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 22, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The wjbmattingly/nara-qwen-3.5-2b is a 2.3 billion parameter vision-language model, fine-tuned by wjbmattingly from Qwen3.5-2B, specifically for cultural heritage transcription. It excels at accurately transcribing English-language archival documents, such as Revolutionary War pension files, including both printed forms and 18th-19th century American cursive. This model is optimized for reading historical documents line-by-line, integrating printed and handwritten text, and achieves a micro CER of 0.2688 on its specialized dataset.

Loading preview...

Overview

wjbmattingly/nara-qwen-3.5-2b is a 2.3 billion parameter vision-language model, fine-tuned from Qwen/Qwen3.5-2B, specifically for cultural heritage transcription of English-language archival documents. This model is the result of extensive fine-tuning on 38,756 pages from 1,407 documents of the United States National Archives Revolutionary War pension files, encompassing printed forms, catalogue cards, and 18th- and 19th-century American cursive.

Key Capabilities

  • Specialized Transcription: Designed to transcribe historical English documents, reading printed text and handwriting together line-by-line.
  • High Accuracy on NARA Data: Achieves a micro Character Error Rate (CER) of 0.2688 and a Word Error Rate (WER) of 0.3205 on a held-out test set of NARA documents.
  • Robust to Historical Scripts: Successfully retrained from a medieval Latin model, overcoming previous issues with historical English orthography (e.g., transcribing 'SERVICE' correctly instead of 'SERUICE').
  • Image-to-Text: Processes page images and returns corresponding text, with a visual token budget of 2,048 per page.

Good For

  • Cultural Heritage and Digital Humanities: Ideal for making handwritten archival material searchable and readable at scale.
  • Research Artifacts: Suitable for academic research involving historical document analysis and transcription.
  • Specific Archival Collections: Particularly effective for US Revolutionary War pension files and similar English-language historical documents.

Important Considerations

  • Domain Specificity: This model is highly specialized. Performance will degrade on other hands, languages, document genres, or centuries. It is not a general English HTR solution.
  • Error Distribution: Errors are not evenly spread; sparse catalogue cards may show higher error rates than dense prose. Users should consider micro vs. macro CER based on their material.
  • Human Verification: Output should not be treated as authoritative without human verification, especially for legal or genealogical records, due to potential errors in names or dates.

Finetunes
1 models