kraken-german-xix-v1

Handwritten-text-recognition model trained with training-atr-models. These are the weights of the best validation checkpoint of the run below — not its last epoch.

Evaluation

Measured on Federal Council minutes 1848-1903 (doi:10.5281/zenodo.4746342), 2,751 lines (held-out documents not in training). The split score below is this run's own validation score; what that split was is in Training data per corpus, and it is a different set from this benchmark either way.

Training data

• dh-unibe/image-text_zh-regierungsratsprotokolle
  • Training projects: MM_1_001, MM_1_002, MM_1_003, MM_1_004, MM_1_005, MM_1_006, MM_1_007, MM_1_008, MM_1_009, MM_1_010, MM_2_101, MM_2_102, MM_2_103, MM_2_104, MM_2_105, MM_2_106, MM_2_107, MM_2_108, MM_2_109, MM_2_110
  • Evaluation: held-out projects MM_1_011, MM_1_012, MM_2_111, MM_2_112
  • Page cap: 8000
• dh-unibe/image-text_kurrent-xix
  • Training projects: TRAIN_CITlab_Bassermann_0_4, TRAIN_CITlab_Bassermann_Manuscripts, TRAIN_CITlab_Bassermann_Manuscripts_0_2, TRAIN_CITlab_Binder_Kochbuch_2, TRAIN_CITlab_Escher_M1, TRAIN_CITlab_Gusbeth, TRAIN_CITlab_Handschriftliche_Archivquellen, TRAIN_CITlab_Konzilsprotkolle_B_Schwartz_5_2018, TRAIN_CITlab_MargareteSick, TRAIN_CITlab_MargareteSick20180731, TRAIN_CITlab_MargareteSick20180731a, TRAIN_CITlab_MÜLLER, TRAIN_CITlab_Protokoll_Hoftheater_1806, TRAIN_CITlab_Rehlen_1834_a, TRAIN_CITlab_RvE_Barlaam_HS_D_, TRAIN_CITlab_Steiner, TRAIN_CITlab_Suppes, TRAIN_CITlab_Suppes_3500, TRAIN_CITlab_Tagebuch_Arnold_v1, TRAIN_CITlab_umkc_Roland_M1, TRAIN_CITlab_umkc_Roland_M2, hufeland_privatbesitz_1829, nn_msgermqu2124_1827, nn_msgermqu2345_1827, parthey
  • Evaluation: held-out projects TEST_CITlab_Bassermann_0_4, TEST_CITlab_Bassermann_Manuscripts, TEST_CITlab_Bassermann_Manuscripts_0_2, TEST_CITlab_Binder_Kochbuch_2, TEST_CITlab_Escher_M1, TEST_CITlab_Gusbeth, TEST_CITlab_Handschriftliche_Archivquellen, TEST_CITlab_Konzilsprotkolle_B_Schwartz_5_2018, TEST_CITlab_MargareteSick, TEST_CITlab_MargareteSick20180731, TEST_CITlab_MargareteSick20180731a, TEST_CITlab_MÜLLER, TEST_CITlab_Protokoll_Hoftheater_1806, TEST_CITlab_Rehlen_1834_a, TEST_CITlab_RvE_Barlaam_HS_D_, TEST_CITlab_Steiner, TEST_CITlab_Suppes, TEST_CITlab_Suppes_3500, TEST_CITlab_Tagebuch_Arnold_v1, TEST_CITlab_umkc_Roland_M1, TEST_CITlab_umkc_Roland_M2
  • Page cap: 19808
• dh-unibe/image-text_parlamentsdienste-protokolle
  • Training projects: 1849_01, 1849_02, 1849_03, 1849_04, 1849_05, 1849_06, 1849_07, 1849_08, 637157213248534270_verkleinert
  • Evaluation: a seeded page-level split of the training projects (partition=0.9, seed=42)
  • Page cap: 138
• dh-unibe/image-text_nr-sr-vereinigte-bundesversammlung-xix
  • Training projects: Nationalrat_01__Sitzung_–_06_11_1848, Nationalrat_02__Sitzung_–_07_11_1848, Nationalrat_03__Sitzung_–_08_11_1848, Nationalrat_04__Sitzung_–_09_11_1848, Nationalrat_05__Sitzung_–_10_11_1848, Nationalrat_06__Sitzung_–_13_11_1848, Nationalrat_07__Sitzung_–_14_11_1848, Nationalrat_08__Sitzung_–_15_11_1848, Nationalrat_09__Sitzung_–_18_11_1848, Nationalrat_10__Sitzung_–_21_11_1848, Nationalrat_11__Sitzung_–_22_11_1848, Nationalrat_12__Sitzung_–_23_11_1848, Nationalrat_13__Sitzung_–_24_11_1848, Nationalrat_14__Sitzung_–_25_11_1848, Nationalrat_15__Sitzung_–_27_11_1848, Nationalrat_16__Sitzung_–_28_11_1848, Nationalrat_17__Sitzung_–_29_11_1848, Ständerat_01__Sitzung_–_06_11_1848, Ständerat_02__Sitzung_–_08_11_1848, Ständerat_03__Sitzung_–_09_11_1848, Ständerat_04__Sitzung_–_11_11_1848, Ständerat_05__Sitzung_–_14_11_1848, Ständerat_06__Sitzung_–_15_11_1848, Ständerat_07__Sitzung_–_21_11_1848, Ständerat_08__Sitzung_–_23_11_1848, Ständerat_09__Sitzung_–_24_11_1848, Ständerat_10__Sitzung_–_25_11_1848, Ständerat_11__Sitzung_–_27_11_1848, Ständerat_12__Sitzung_–_28_11_1848, Ständerat_13__Sitzung_–_28_11_1848, VBV_01__Sitzung_–_16_11_1848, VBV_02__Sitzung_–_17_11_1848, VBV_03__Sitzung_–_20_11_1848, VBV_04__Sitzung_–_24_11_1848, VBV_05__Sitzung_–_29_11_1848
  • Evaluation: a seeded page-level split of the training projects (partition=0.9, seed=42)
  • Page cap: 52

Materialized from that selection: 15,527 pages, 405,955 transcribed lines.

Hyperparameters

spec: '[256,64,0,1 Cr4,2,8,4,2 Cr4,2,32,1,1 Mp4,2,4,2 Cr3,3,64,1,1 Mp1,2,1,2 S1(1x0)1,3
  Lbx256 Do0.5 Lbx256 Do0.5 Lbx256 Do0.5 Cr255,1,85,1,1]'
batch_size: 32
schedule: cosine
lrate: 0.0001
quit: early
epochs: 90
min_epochs: null
lag: 10
augment: true
normalization: NFD
normalize_whitespace: true
weights_format: coreml
resize: union
freeze_backbone: null
accumulate_grad_batches: 1
pad: null
warmup: null
seed: 42
workers: 8
device: cuda:0
benchmarks:
- hf_repo: dh-unibe/image-text_federal-minutes-testset
  project: TEST_federal_minutes
  split: train
  label: Federal Council minutes 1848-1903 (doi:10.5281/zenodo.4746342), 2,751 lines

Provenance

metadata.json in this repo is the record the trainer wrote, verbatim: the full request, the parsed metrics and the job id.

Using it

pip install kraken
python -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('dh-unibe/kraken-german-xix-v1', 'kraken-german-xix-v1.mlmodel'))"
kraken -i page.jpg page.txt segment -bl ocr -m <the printed path>

Notes

#164, run 1: the first kraken model on the XIX corpora, against qwen3vl-german-xix-v2 (7.65 %) and qwen3.5-4b-german-xix-v2 (6.80 %) on the same benchmark. Fine-tuned from kraken-bohemian_19th, the only registry base that is Kurrent AND 19th century AND German, and height 120 — which this material needs: the kraken+ spec of all five earlier runs leaves 1.29-1.63 CTC frames per character here, height 120 gives 2.42-3.05 (389328f). zh-regierungsratsprotokolle is held out by VOLUME rather than by page, and its three Testset_/TEST_CITlab_ directories are excluded from training.

Correction, 08.10.2026 (training-atr-models 4de9468). The sentence above about "the kraken+ spec of all five earlier runs" is wrong, and it was written into this request before it was checked. ketos train ignores --spec when --load is given, so four of those five runs inherited their architecture from their base, and every base this project has fine-tuned from is height 120 already. The height-64 spec was used once, by the only from-scratch run, which is also the worst by a factor of three. Height 120 remains the right choice for this material (2.42-3.05 CTC frames per character against 1.29-1.63) — but it came with the base, not as a correction to a practice that did not exist. The job record keeps the request as submitted; this is the publication record.