Kurrent Transcriber 1B

GGUF for llama.cpp: 13fragments/kurrent-transcriber-1b-GGUF

Try it in your browser: demo Space. The demo runs on transformers, so its text can differ slightly from the results below.

Kurrent Transcriber 1B transcribes pages of historical German handwriting. It reads Kurrent, Sütterlin, and related hands. You give it a page image. It returns the page text, line by line, as written.

Its output is a draft. Check every line against the scan before you use it.

  • What it is: a weight merge of two fine-tunes of LightOnOCR-3-1B: our own fine-tune (not published) and LightOnOCR-HGH by Bernhard Straub. Its vision encoder and vision projection are the average of the two; its language model is the base model's.
  • Base model: LightOnOCR-3-1B (Apache-2.0).
  • Weights licence: CC BY-SA 4.0.
  • Inference code: inference/ (Apache-2.0). You need this code to get the reported results.
  • GGUF for llama.cpp: see the link above. The GGUF files have no benchmark result.
  • Version: v1, 2026-11. Archived on Zenodo: 10.5281/zenodo.24342061 (all versions).

Results in short

On the 53 held-out pages of the Kurrent Pages Benchmark, in one run:

  • Character error rate: level with Transkribus German Genius, the release gate. Its median normalized character error rate (cer_norm) is 0.071. German Genius has 0.131. This gate tests "not worse"; a claim of "better" was not registered. See Results on the held-out benchmark.
  • Against its parents: in this run, on these 53 pages, it reads better than LightOnOCR-HGH (0.090) and our own fine-tune (0.112). Neither interval covers changes between serving sessions. A llama.cpp development check of this lead failed its limit, set before the check, and was waived. See Known weaknesses.
  • Names, places, and dates: level with LightOnOCR-HGH and slightly behind German Genius. Median entity recall: 0.83, against 0.85 for German Genius.
  • Invented text: none found. Two AI model reviewers found no invented passage on the 53 pages.
  • Frontier models: in single runs on one date each, two general-purpose frontier models had median cer_norm 0.055 and 0.070.
  • A development gate failed and was waived. The image step can remove library stamps written next to colour charts. See How this model was chosen.

On three public test sets from other groups, it reads better than LightOnOCR-HGH on all three. The details are in Results on outside benchmarks.

Intended use

  • Draft transcription of Kurrent and Sütterlin pages for a person who checks it against the scan.
  • Offline use on your own hardware. It has about 1 billion parameters.
  • Whole pages, with no line segmentation step. If you use a line model such as TrOCR-Kurrent, give this model the page image, not line crops.

Out of scope

  • Use without a person who checks the text.
  • Extraction of names, dates, or places without review.
  • Legal, genealogical, or archival evidence that rests on the transcription alone.
  • Line images: the model reads whole pages, and on a line crop it also writes the neighbouring lines.
  • Languages other than German: almost all training data is German.

How to run

Weights alone do not give the reported results. Four fixed parts make up the release engine, and the client in inference/ applies all of them:

  1. Image step. Before reading, the client paints over the furniture outside the leaf: colour charts, rulers, library footer bands, and the margins around the leaf. Marks on the leaf stay.
  2. Request. The page image alone, without a text prompt. Greedy decoding, at most 4,096 new tokens, and a windowed repetition penalty of 1.1 over the last 20 tokens. vLLM needs a server plugin for the penalty.
  3. Output map. The model writes some markup: for struck, inserted, and underlined text, for an illegible spot, and for table cells. The client removes the marks and keeps the text, writes [unleserlich] for an illegible spot, and writes | between table cells. Its JSON output also keeps the unmapped reply.
  4. Output step. A reply that reaches 4,096 tokens is a repetition loop. The client cuts the loop and marks the page incomplete: "repetition". In any other reply, exact repeated runs are removed.

Download this repository, then serve the model with vLLM on an NVIDIA GPU:

hf download 13fragments/kurrent-transcriber-1b --local-dir kurrent-transcriber-1b
cd kurrent-transcriber-1b
python3.11 -m venv serve
serve/bin/pip install -r inference/server/requirements.txt
PYTHONPATH=inference/server VLLM_USE_FLASHINFER_SAMPLER=0 serve/bin/vllm serve . \
  --served-model-name kurrent-transcriber-1b --max-model-len 16384 --dtype bfloat16 \
  --limit-mm-per-prompt '{"image":1}' \
  --logits-processors windowed_repetition_penalty:WindowedRepetitionPenalty

pip install huggingface_hub installs the hf command.

Then install the client in its own environment and transcribe pages (JPEG, PNG, or WebP):

python3.11 -m venv client
client/bin/pip install ./inference
client/bin/kurrent-transcribe page.jpg
client/bin/kurrent-transcribe --out transcripts/ scans/*.jpg

inference/README.md describes the output fields and the fixed settings. The results below were measured with vLLM 0.23.0, transformers 5.13.0, and one NVIDIA RTX PRO 6000 GPU. Other runtimes give different output. Transformers alone is not measured. llama.cpp was measured on development pages only (see Known weaknesses).

Results on the held-out benchmark

Kurrent Pages Benchmark v1.0, ground truth v1.0, 53 scored pages, one sealed run on 2026-11-17. The leaderboard lists this model since benchmark release v1.0.4.

Engine Run cer_norm median Hard strata (33 pages) cer_norm median Entity recall median
Kurrent Transcriber 1B 2026-11-17 0.071 0.091 0.83
LightOnOCR-HGH (parent; no image step) 2026-11-17 0.090 0.098 0.83
Our own fine-tune (parent, not published) 2026-11-13 0.112 0.122 0.75
Transkribus German Genius (commercial reference) 2026-09-24 0.131 0.128 0.85
Transkribus Text Titan II (commercial reference) 2026-09-24 0.169 0.150 0.857
CHURRO 3B (open, zero-shot) leaderboard v1.0 0.258 0.266 0.588
LightOnOCR-3-1B (base model, its card's settings) 2026-11-19 0.359 0.363 0.500

This model and LightOnOCR-HGH ran in the same session, with the same request settings and output map; LightOnOCR-HGH ran without the image step and run removal. German Genius and our own fine-tune come from earlier runs; their stored outputs were scored again under the same ground truth for the paired comparisons below. The last three rows come from the benchmark leaderboard. The base model ran in another session with the settings of its own model card (one sample at temperature 0.2). Its row shows the model before fine-tuning; it is not a paired comparison. A planned CHURRO 3B run in the same session did not start (a launcher error), so it has no row of its own.

Frontier snapshots. These rows are single runs on one date, through the vendor's command-line tool, with the benchmark prompt and the page image. Vendor models change. They are not ranked with the table above on the leaderboard.

Engine Run cer_norm median Hard strata (33 pages) cer_norm median Entity recall median
Claude Fable 5.1 (effort max) 2026-10-31 0.055 0.058 0.944
GPT-6 Astra (reasoning effort max) 2026-11-06 0.070 0.106 0.933

Paired comparisons. The difference in cer_norm per page, this model minus the other engine, with the mean over the 53 pages and a 95 % page bootstrap interval:

Against Mean difference 95 % interval Median difference
Transkribus German Genius −0.073 [−0.116, −0.038] −0.020
LightOnOCR-HGH −0.021 [−0.045, −0.001] −0.008
Our own fine-tune −0.049 [−0.082, −0.019] −0.014

In this run, on these 53 pages, the model passes the registered "level with German Genius" gate: the interval must end below +0.03, and it ends at −0.038. This gate does not test "better", and the card makes no such claim. All intervals cover the choice of pages only. They do not cover changes between serving sessions. In an earlier pair of sessions on public pages, LightOnOCR-HGH reproduced its text on 67 of 274 pages. Against LightOnOCR-HGH, the lead is small: the upper end of its interval is −0.001.

Loops. On none of the 53 pages did the reply reach the token limit.

How this model was chosen

Read the results with these conditions:

  • Chosen on other pages. We chose the merge on development pages and on public page sets, not on the benchmark. No benchmark output of the merge existed before this run.
  • A development gate was waived. Because the image step paints over colour charts, their printed text is not transcribed. On development pages with library stamps, the model had written some stamps in the same text block, and these were lost too. Stamp text appeared on 9 stamp pages with the image step and on 17 without it. Passing the gate required at least 11. 13fragments waived the gate after this result and states it here. Stamps written elsewhere on the page are not affected.
  • Benchmark pages seen before by a parent. Our own fine-tune ran twice on the benchmark pages (2026-11-09 and 2026-11-13). Only aggregate scores and the review procedure were read. Half of the merge's vision weights come from it. By the benchmark rule, one benchmark run per training-data round is allowed; this run is a stated exception, because the merge adds no training data.
  • Training sources of LightOnOCR-HGH. None of its ten training sources is among the benchmark's source collections (a check by collection name and record link).

Invented text

Two AI model reviewers checked the model's output on all 53 benchmark pages against the page images and the reference transcriptions: Claude Fable 5.1 and GPT-6 Astra, each at maximum effort. A passage counts as invented when either reviewer marks it. No human reviewed these pages.

Measure Result Release gate
Median share of invented text per page 0.000 at most 0.01: pass
Pages with more than 0.10 invented text 0 none: pass
Pages with any invented text 0 of 53 —

Each reviewer alone also found no invented passage. LightOnOCR-HGH, reviewed the same way in the same run, also had no page with invented text.

Development pages. An automatic flag for possible invented text marked more development pages for the merge than for either parent on one letter collection (21 pages, against 16 for LightOnOCR-HGH and 11 for our own fine-tune). These flags were not reviewed.

What this means for you: no invented passage was found on these 53 pages. A model of this kind can still write a plausible passage that is not on the page. Check every line against the scan, also lines that read well.

Results on outside benchmarks

These public test sets come from other groups. They are outside our own fine-tuning data. We did not check the training sources of LightOnOCR-HGH and of the base model against these pages. Four engines ran in one session on 2026-11-16, on one GPU and one serving stack.

Each score is the CHURRO-DS metric, normalized Levenshtein similarity (NLS): the mean over pages, times 100, higher is better. Mean cer_norm (lower is better) is in brackets.

Test set Pages Kurrent Transcriber 1B LightOnOCR-HGH Our own fine-tune
ICFHR 2016 READ competition, Ratsprotokolle, test set 50 71.6 (0.296) 68.2 (0.335) 69.4 (0.320)
BullingerDB, unseen-writer test split 200 64.3 (0.454) 62.2 (0.465) 60.4 (0.492)
CHURRO-DS test, German handwriting 24 76.4 (0.626) 70.1 (0.652) 73.4 (0.648)

Paired NLS differences, this model minus LightOnOCR-HGH, with 95 % bootstrap intervals: ICFHR 2016 +3.3 [+1.9, +4.9], Bullinger +2.1 [+1.4, +2.9], CHURRO-DS +6.3 [+2.4, +11.1]. These are nine comparisons in all, without a correction for multiple tests.

CHURRO 3B, an open 3B model for historical documents, ran in an earlier session (2026-11-15) with the settings of its own benchmark code. Unpaired mean NLS: ICFHR 2016 81.5, Bullinger 62.1. CHURRO 3B was trained on other pages of the ICFHR 2016 collection. On CHURRO-DS, no comparison is made: a reference row moved by more than 1.0 point between the sessions.

Known weaknesses

  • Stamps near colour charts. See How this model was chosen.
  • Names, places, and dates. Entity recall is 0.83 against 0.85 for German Genius. Names are often misread.
  • Invented text. See Invented text.
  • Loops. A reply can loop until the token limit. Then the client cuts the loop, and the page text after the loop start is missing. Such pages carry incomplete: "repetition".
  • Whole-page repeats. Sometimes the model writes the whole page twice. Run removal does not catch this, because it removes only repeated runs of 20 to 400 characters. On 450 Wikisource development pages that neither parent trained on, read with vLLM, 2 replies were more than 1.6 times as long as the transcription. LightOnOCR-HGH had 1. Check a long reply against the scan.
  • llama.cpp check failed. On 40 of these development pages, read with llama.cpp (vision weights bf16, text weights q8_0), the mean cer_norm difference against LightOnOCR-HGH was +0.019, 95 % interval [−0.023, +0.087]. The median difference was −0.006. One page that the model wrote twice gives most of the mean. Set before the check, the limit was an upper end below +0.02, so the check failed. 13fragments waived it for its reading app and states it here. On all 450 pages with vLLM, after the check and for description only: mean −0.018 [−0.053, +0.003], median −0.003.
  • Two-page spreads. On a wide spread, the model can leave out a block of text. Split spreads into single pages first.
  • Furniture on the leaf. The image step does not remove colour charts that lie on the leaf, grey charts, or rulers without colour patches. The model can transcribe their print.
  • Line images. On a line crop, the model also writes the neighbouring lines, so give it whole pages.
  • Serving changes. Output changes slightly with the serving stack. Comparisons are exact only in one serving state.

Training

The merge itself trains nothing. It averages two fine-tunes of the same base model revision.

  • Merge rule. The 222 tensors of the vision encoder and the vision projection are the mean of the two parents, computed in float32 and stored in bfloat16. All 310 language-model tensors are equal in both parents and the base model, and are kept unchanged.
  • Parent A, our own fine-tune. Base LightOnOCR-3-1B; vision encoder and vision projection trained in full (409.6 million parameters); 2,400 steps, learning rate 2e-4, gradient clip 0.5, seed 1702, maximum sequence length 8,192, page image only; 36,818 training rows from the sources in the first table below. No benchmark page and no development page is in its training. No frontier-model output is in its training data.
  • Parent B, LightOnOCR-HGH by Bernhard Straub, revision 47931fbe, CC BY-SA 4.0. Base LightOnOCR-3-1B; vision encoder and vision projection trained; 21,038 training pages from the sources in the second table below. See its model card for the recipe. The datasets field of this card lists the three of its sources that are on Hugging Face.

Training data of parent A

Source Rows Licence
ICDAR 2017 READ, Train-A and Train-B, German pages 7,875 CC BY 4.0
Senatsprotokolle, University of Tübingen (19th-century Kurrent) 7,913 CC BY 4.0, with written permission
Dresdner Hofdiarium 1653–56 and 1673 (Kurrent) 716 CC BY 4.0
ScaDS.AI German line and word handwriting 4,000 CC BY 4.0
Deutscher Reichsanzeiger ground truth (Fraktur print) 4,000 CC0 1.0
Kosmos lecture notes, Deutsches Textarchiv 1,812 CC BY 4.0
Letters of Fanny Mendelssohn to Wilhelm Hensel 102 CC BY 4.0
Akademieprotokolle, Austrian Academy of Sciences 1,253 CC BY 4.0, with written permission
Konzilsprotokolle, University of Greifswald 173 CC BY 4.0
Synthetic Kurrent pages (generated) 4,989 generated; fonts SIL OFL 1.1
Synthetic Fraktur pages (generated) 2,985 generated from CC0 text; fonts SIL OFL 1.1
Synthetic blank and illegible pages (generated) 1,000 generated from CC0 text; fonts SIL OFL 1.1

Training data of parent B (from the LightOnOCR-HGH model card, revision 47931fbe)

Source Holder Pages Text licence Images
Digitale Briefedition Alfred Escher (Zenodo 10927129) Staatsarchiv des Kantons Zürich 13,156 CC BY-SA 4.0¹ CC BY-SA 4.0¹
Kurrent XIX, Regierungsratsprotokolle Staatsarchiv Zürich / DH Uni Bern 1,796 MIT MIT
Historisches Grundbuch Basel DH Uni Bern 1,907 MIT MIT
DTA, Hidden Kosmos Deutsches Textarchiv (BBAW); images SBB 1,640 CC BY 4.0 Public Domain Mark (SBB)
edition humboldt digital (v12.0.0) BBAW 925 CC BY-SA 4.0¹ Public Domain Mark (SBB)
Nägeli-Briefe Zentralbibliothek Zürich 680 CC BY-SA 4.0 Public Domain Mark (e-manuscripta)
Frauen im Fokus SBB / Universität Potsdam 180 CC0 1.0 Public Domain Mark (SBB)
Fanny Hensel SBB / UdK Berlin 158 CC BY 4.0 Public Domain Mark (SBB)
German 20th century (3 subsets) DH Uni Bern 584 MIT MIT
Gantner (DTA part) M. Gantner 12 CC BY 4.0 CC BY 4.0 (DTA part)

¹ The LightOnOCR-HGH card states: licence clarification in process, a move to CC BY 4.0 without ShareAlike.

The MIT notice of the three dh-unibe datasets is in NOTICE.

Credits

Both parents were trained on the work of the creators and institutions below. Their material was adapted: page images and transcriptions were converted into training examples, and model weights were trained on them. No source endorses this model. training_data.json lists each source with its licence and link.

Parent A, our own fine-tune:

Parent B, LightOnOCR-HGH:

Models:

Licence

  • Model weights: CC BY-SA 4.0. If you share adapted weights, share them under the same licence.
  • Base model: LightOnOCR-3-1B, Apache License 2.0. See NOTICE.
  • Inference code: Apache License 2.0.

Citation

@misc{kurrenttranscriber1b_2026,
  title        = {Kurrent Transcriber 1B: a page transcription model for historical German handwriting},
  author       = {{13fragments}},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.24342061},
  howpublished = {\url{https://huggingface.co/13fragments/kurrent-transcriber-1b}}
}

Please also cite the parent model and the base model:

@misc{straub2026historichtr,
  title        = {LightOnOCR-3-1B · Historic German Handwriting},
  author       = {Straub, Bernhard},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/b-straub/LightOnOCR-3-1B-Historic-German-Handwriting}}
}

@misc{lightonocr2_2026,
  title        = {LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR},
  author       = {Said Taghadouini and Adrien Cavaill\`{e}s and Baptiste Aubertin},
  year         = {2026},
  howpublished = {\url{https://arxiv.org/abs/2601.14251}}
}
Downloads last month
55
Safetensors
Model size
1B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for 13fragments/kurrent-transcriber-1b

Datasets used to train 13fragments/kurrent-transcriber-1b

Space using 13fragments/kurrent-transcriber-1b 1

Collection including 13fragments/kurrent-transcriber-1b

Paper for 13fragments/kurrent-transcriber-1b