Qwen3-VL-4B Trithemius OCR
Qwen3-VL-4B Trithemius is a specialized vision-language model for reading and transcribing early printed works by Johannes Trithemius (1462–1516). Built on Qwen3-VL-4B-Instruct, it was fine-tuned using page-image and transcription pairs derived from public-domain scans held by libraries and other cultural institutions.
The model is intended to produce a diplomatic, structure-preserving transcription of difficult early print: blackletter and worn roman type, long s, ligatures, abbreviations, running heads, marginal material, names, dates, and mixed tabular or enumerative pages.
235B teacher distillation
Training used OCR labels generated or adjudicated by
Qwen3-VL-235B-A22B-Instruct as a teacher for the 4B student. In the project’s page-level OCR comparison, the resulting 4B Trithemius model
matched the 235B teacher on 86% of evaluated pages.
This is the most important result of the fine-tune: within its narrow early-print OCR domain, a 4B model recovered most of the teacher’s useful transcription behavior while remaining small enough to quantize and run locally. The 86% figure is a project-specific OCR agreement rate—not character error rate, not a claim of general vision-language parity, and not evidence that the 4B model matches the 235B teacher outside this historical-OCR task.
Files in this repository
| File | Purpose |
|---|
Qwen3-VL-4B-Trithemius-Q6_K.gguf | Q6_K quantized language-model weights for llama.cpp-compatible runtimes. |
mmproj-Qwen3-VL-4B-Trithemius-F16.gguf | Required F16 vision projector. |
The Q6_K model and projector must be used together. Full model tensors are being added separately so that the fine-tune can also be used and converted outside the GGUF runtime.
Recommended use
This is an OCR/transcription model, not an English translation model. Supply one page image at a time and request transcription rather than commentary, modernization, or reconstruction.
The production Trithemius run used llama.cpp-compatible vision inference with settings equivalent to:
1--model Qwen3-VL-4B-Trithemius-Q6_K.gguf
2--mmproj mmproj-Qwen3-VL-4B-Trithemius-F16.gguf
3--no-mmproj-offload
4--ctx-size 8192
5--image-min-tokens 1024
6--max-tokens 3000
Example instruction:
1Transcribe this historical printed page faithfully. Preserve line and paragraph
2structure, headings, names, dates, abbreviations, and visible marginal or
3tabular material. Do not translate or silently repair uncertain text. Mark text
4that cannot be read reliably as [unclear]. Return only the transcription.
For dense pages, splitting the image into overlapping regions is safer than simply increasing the generation limit. A larger token cap can extend a repetition loop instead of recovering missing text.
What shipped from this model
The model supplied the Latin OCR witness for 27 Trithemius 4B editions published in the Trithemius Corpus. Those are the project’s recommended English reading editions where available. Their English was produced separately with GPT-5.5 dual-context translation, independently assessed by Claude Sonnet 5, and then reviewed and repaired at full-work level.
The model card deliberately separates OCR quality from downstream translation quality. An early corpus report attributed too much failure to the LoRA OCR. Direct inspection of 3,902 pages from eight difficult works found body-page OCR to be generally clean, with total degenerate-page rates of approximately 0.2–3.3% per audited work. The larger failure cluster came from an unguarded downstream translation run that could drift to neighboring pages. That translation failure is not a model benchmark for this OCR model.
See the corpus
Methodology for the complete OCR, translation, grading, and remediation record.
Strengths
- Early printed Latin in blackletter and worn roman type
- Long-s and common ligature handling
- Preservation of headings, lineation, lists, and page structure
- Better recovery of names and abbreviated forms than embedded PDF text or ordinary Tesseract on the target material
- Honest transcription workflow when prompted to use
[unclear]
Limitations
- This is a domain fine-tune, not a general-purpose OCR benchmark winner.
- Dense catalogues, cipher alphabets, tables, Greek, Hebrew, marginalia, stamps, and severely damaged leaves remain difficult.
- The model can repeat lines or enumerative content, especially on degraded or token-capped pages.
- It can occasionally generate plausible-looking letter sequences where the page is unreadable. Always compare consequential readings with the facsimile.
- Approximately 58 pages in the production campaign reached the original 3,000-token cap; some were truncated and some entered repetition loops.
- The audit figures above describe degeneration screening on a project corpus, not a controlled CER/WER evaluation against a fully human-corrected benchmark.
- The published T4B translations include downstream translation and editorial work and should not be treated as raw demonstrations of model-only output.
Intended users
Digital-humanities researchers, historians of the book, Latinists, librarians, and OCR practitioners working with fifteenth- and sixteenth-century European print. Outputs are suitable as machine witnesses and research starting points, not as critical editions without page-image verification.
Training data and provenance
Training examples were derived from public-domain institutional page scans and transcription pairs assembled during the Trithemius Corpus project. The project source code, methodology, work metadata, provenance records, and published OCR witnesses are available in the
Trithemius Corpus repository.
The training pipeline accepted conservative multi-witness labels and supported teacher adjudication for pages on which weaker OCR witnesses disagreed. The 235B teacher was used to turn this evidence into grounded page transcriptions for the 4B student; expansive, translated, repetitive, or visibly unsupported teacher output was not intended to be treated as gold.
The released model should not be used to infer that libraries or scan providers endorsed the fine-tune. Source institutions retain attribution for their digitized holdings; the underlying historical works are public domain.
License
Released under the Apache License 2.0, consistent with the Qwen base model. Users remain responsible for complying with licenses or terms attached to any images they process.
Citation
1Fabin, Ian Carlos (2026). Qwen3-VL-4B Trithemius OCR. Hugging Face model repository.
2https://huggingface.co/Carlosian/trithemius-q6-models
Author
Developed and assembled by
Ian Carlos Fabin (
Carlosian).
