OCR/HTR Models for Postwar Japanese Student and Social Movement Documents

Abstract

This release provides trained OCR/HTR models and supporting software for the automated transcription of postwar Japanese student and social movement documents held by Tohoku University Archives. The released components were selected from the configurations evaluated in the associated research and are intended to support further research and practical experimentation with historical document transcription. Model components and contents. The package combines a fine-tuned NDL-OCR-based text-line detector with a LoRA-adapted text recognizer based on Qwen3.6-35B-A3B. Together with NDL-OCR reading-order processing, these components support the extraction and recognition of text from document images. The release includes trained model weights, inference scripts, configuration files, installation instructions, license notices, and reproducibility documentation. The individual model checkpoints and their correspondence to the experimental configurations are documented in the accompanying technical files. Original archival images, reference transcriptions, and upstream foundation-model weights are not included. Validation and reproducibility. The packaged inference pipeline was tested on a document page containing 49 text lines. The resulting transcription and reading-order XML matched previously archived outputs byte-for-byte. All 26 locally cached Qwen base-model weight shards were verified against the corresponding upstream SHA-256 checksums. These checks establish consistency with previously obtained outputs but do not constitute an independent evaluation of transcription accuracy against reference transcriptions. Installation and execution on a separate machine have not yet been independently verified. Download and reconstruction. Because of upload transfer limitations, the original 342,528,241-byte ZIP archive is distributed in 11 numbered parts (.part-001 to .part-011). Download all 11 parts and combine them in numerical order according to the accompanying README_REASSEMBLE.txt. The file PART_SHA256SUMS.txt provides checksums for the individual parts. The reconstructed ZIP archive must have the following SHA-256 checksum: 3adb536083eee5068444b74f163c4812ab238799ace77a8cd10883cd53e63424 Related dataset. The underlying collection of historical document images, layout annotations, and reference transcriptions is available separately on Zenodo: Version 1.0.0: https://doi.org/10.5281/zenodo.18592517 Version 1.1.0: https://doi.org/10.5281/zenodo.21798820 The underlying dataset and the present model release are distinct research outputs. The source dataset does not provide all the reading-order information used in the associated research. Intended use and limitations. The models are provided for research, experimental OCR/HTR processing, and preliminary transcription of historical Japanese documents. Recognition accuracy may vary with handwriting styles, printing methods, document layouts, and image quality. Automatically generated transcriptions should be checked against the original document images before scholarly quotation or use as authoritative historical evidence. The release does not guarantee accurate transcription of arbitrary document collections. Licenses and attribution. The package contains components subject to different licenses. Please consult LICENSE_MAP.md, NOTICE.md, and MODEL_CARD.md in the reconstructed archive for component-specific licensing, attribution, technical specifications, and usage information.