Loading…
Loading…
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.
| metric | value |
|---|---|
| split CER | 8.89 % |
| split WER | 24.71 % |
| samples scored | 500 |
| characters scored | 23736 |
| character errors | 2109 |
Measured on this run's own held-out validation split (page-level and seeded, so no page contributes lines to both sides). It is not a score on a shared benchmark and does not transfer to a different corpus.
The validation split is a seeded partition of the training projects. It measures the model on pages it has not seen, in hands it has. Expect a higher error rate on a new writer.
dh-unibe/image-text_zh-regierungsratsprotokolle @ f6a06745bb134044d5f937b22000128cc4bc24fdpartition=0.9, seed=42)dh-unibe/image-text_kurrent-xix @ 20743e9f6982460b243e014091626da43e2b57f7MM_2_033, MM_2_034, MM_2_035, MM_2_036, MM_2_037, MM_2_038, MM_2_039, MM_2_040, MM_2_041, MM_2_042, MM_2_043, MM_2_044, MM_2_045, MM_2_046, MM_2_047, MM_2_048, MM_2_049, MM_2_050, TEST_CITlab_Bassermann_0_4, TEST_CITlab_Bassermann_Manuscripts, TEST_CITlab_Bassermann_Manuscripts_0_2, TEST_CITlab_Binder_Kochbuch_2, TEST_CITlab_Binder_Kochbuch_M3+, TEST_CITlab_DreierM2, TEST_CITlab_Escher_M1, TEST_CITlab_GemeindeChronikM1, TEST_CITlab_GrimmBriefe, TEST_CITlab_Gusbeth, TEST_CITlab_Haeckermann_3+, TEST_CITlab_Handschriftliche_Archivquellen, TEST_CITlab_Hoffmann,_Hans, TEST_CITlab_Konzilsprotkolle_B_Schwartz_5_2018, TEST_CITlab_Konzilsprotokolle_A_Haeckermann_2, TEST_CITlab_Konzilsprotokolle_A_Haeckermann_3, TEST_CITlab_Konzilsprotokolle_M4_HTR+, TEST_CITlab_Lotte1, TEST_CITlab_Lotte_Tagebuecher, TEST_CITlab_Lottes_Schrift, TEST_CITlab_MargareteSick, TEST_CITlab_MargareteSick20180731, TEST_CITlab_MargareteSick20180731a, TEST_CITlab_Minutes_of_Estonian_Knighthood, TEST_CITlab_MÜLLER, TEST_CITlab_MÜLLER_A, TEST_CITlab_Nekrolog, TEST_CITlab_Nekrolog_2, TEST_CITlab_Nekrolog_M2, TEST_CITlab_Nekrolog_M3, TEST_CITlab_OEAW_M4+, TEST_CITlab_Protokoll_Hoftheater_1806, TEST_CITlab_Protokoll_Hoftheater_1806-Vers_2018Nov, TEST_CITlab_Protokoll_Hoftheater_1806-Vers_2018Nov+, TEST_CITlab_Pyl_M1, TEST_CITlab_Pyl_M2, TEST_CITlab_Pyl_M3, TEST_CITlab_Pyl_M4, TEST_CITlab_Pyl_M5, TEST_CITlab_Rehlen_1834_a, TEST_CITlab_Roland-Hilde_M1+, TEST_CITlab_Roland_M3+, TEST_CITlab_RvE_Barlaam_HS_D_, TEST_CITlab_Schwartz_M6, TEST_CITlab_St_Aegidi_M1, TEST_CITlab_Steiner, TEST_CITlab_Suppes, TEST_CITlab_Suppes_3500, TEST_CITlab_Tagebuch_Arnold_v1, TEST_CITlab_Todesurteile_1849-1891, TEST_CITlab_Todesurteile_M1+, TEST_CITlab_Tuempel_M3, TEST_CITlab_umkc_Roland_M1, TEST_CITlab_umkc_Roland_M2, TRAINING_TESTSET_Binder_Kochbuch_M3+, TRAINING_TESTSET_Haeckermann_3+, TRAINING_TESTSET_OEAW_M4+, TRAINING_TESTSET_Protokoll_Hoftheater_1806-Vers_2018Nov+, TRAINING_TESTSET_Roland-Hilde_M1+, TRAINING_TESTSET_Roland_M3+, TRAINING_TESTSET_Todesurteile_1849-1891, TRAINING_TESTSET_Todesurteile_M1+, TRAIN_CITlab_Bassermann_0_4, TRAIN_CITlab_Bassermann_Manuscripts, TRAIN_CITlab_Bassermann_Manuscripts_0_2, TRAIN_CITlab_Binder_Kochbuch_2, TRAIN_CITlab_Binder_Kochbuch_M3+, TRAIN_CITlab_DreierM2, TRAIN_CITlab_Escher_M1, TRAIN_CITlab_GemeindeChronikM1, TRAIN_CITlab_GrimmBriefe, TRAIN_CITlab_Gusbeth, TRAIN_CITlab_Haeckermann_3+, TRAIN_CITlab_Handschriftliche_Archivquellen, TRAIN_CITlab_Hoffmann,_Hans, TRAIN_CITlab_Konzilsprotkolle_B_Schwartz_5_2018, TRAIN_CITlab_Konzilsprotokolle_A_Haeckermann_2, TRAIN_CITlab_Konzilsprotokolle_A_Haeckermann_3, TRAIN_CITlab_Konzilsprotokolle_M4_HTR+, TRAIN_CITlab_Lotte1, TRAIN_CITlab_Lotte_Tagebuecher, TRAIN_CITlab_Lottes_Schrift, TRAIN_CITlab_MargareteSick, TRAIN_CITlab_MargareteSick20180731, TRAIN_CITlab_MargareteSick20180731a, TRAIN_CITlab_Minutes_of_Estonian_Knighthood, TRAIN_CITlab_MÜLLER, TRAIN_CITlab_MÜLLER_A, TRAIN_CITlab_Nekrolog, TRAIN_CITlab_Nekrolog_2, TRAIN_CITlab_Nekrolog_M2, TRAIN_CITlab_Nekrolog_M3, TRAIN_CITlab_OEAW_M4+, TRAIN_CITlab_Protokoll_Hoftheater_1806, TRAIN_CITlab_Protokoll_Hoftheater_1806-Vers_2018Nov, TRAIN_CITlab_Protokoll_Hoftheater_1806-Vers_2018Nov+, TRAIN_CITlab_Pyl_M1, TRAIN_CITlab_Pyl_M2, TRAIN_CITlab_Pyl_M3, TRAIN_CITlab_Pyl_M4, TRAIN_CITlab_Pyl_M5, TRAIN_CITlab_Rehlen_1834_a, TRAIN_CITlab_Roland-Hilde_M1+, TRAIN_CITlab_Roland_M3+, TRAIN_CITlab_RvE_Barlaam_HS_D_, TRAIN_CITlab_Schwartz_M6, TRAIN_CITlab_St_Aegidi_M1, TRAIN_CITlab_Steiner, TRAIN_CITlab_Suppes, TRAIN_CITlab_Suppes_3500, TRAIN_CITlab_Tagebuch_Arnold_v1, TRAIN_CITlab_Todesurteile_1849-1891, TRAIN_CITlab_Todesurteile_M1+, TRAIN_CITlab_Tuempel_M3, TRAIN_CITlab_umkc_Roland_M1, TRAIN_CITlab_umkc_Roland_M2, hufeland_privatbesitz_1829, nn_msgermqu2124_1827, nn_msgermqu2345_1827, parthey, semper_20_MSpartition=0.9, seed=42)dh-unibe/image-text_rats-und-richtebuecher_xv-xvi @ 011e82080411a51bcb9e032e9329c919682e5f6cpartition=0.9, seed=42)dh-unibe/image-text_koenigsfelden-charters-post-1500 @ fac683bee395bad466c464bfffd0e6f5299d6c4dpartition=0.9, seed=42)dh-unibe/image-text_bullinger-autoren @ 85ea246a5021a5f244c392f97f2c8ac9a42cc717partition=0.9, seed=42)dh-unibe/image-text_aaeb-xiv-xvii @ 2a9b66116db32c728fd893c3c9af5c41a3fa856dpartition=0.9, seed=42)dh-unibe/image-text_historisches-grundbuch-basel_xix-xx_train @ 897509cae41d53fd43d9f5cd0ad5a0e8079e5893partition=0.9, seed=42)dh-unibe/image-text_german-20th-century @ 47a16a5077a062942237f9817471f60a60993689partition=0.9, seed=42)dh-unibe/image-text_parlamentsdienste-protokolle @ 5702670446f335a54505be3d238a5998cf844a04partition=0.9, seed=42)dh-unibe/image-text_nr-sr-vereinigte-bundesversammlung-xix @ de8ad0f86cf878607eb42390ed637f4e1be70b59partition=0.9, seed=42)dh-unibe/image-text_koenigsfelden-charters-part-2 @ 1d67552f4470c9d9b11ffb09b0b6a104eef185d1partition=0.9, seed=42)dh-unibe/image-text_koenigsfelden-adhr-colmar @ 3cffcc9143da1b8248316a9605bda8680887da67partition=0.9, seed=42)dh-unibe/image-text_aaeb-xiv-xvii-part-2 @ ce2739fce41c0b8f8673127525839842411020b0partition=0.9, seed=42)dh-unibe/image-text_sg-missiven @ 495fc94c5f616d9a9161917021c75e7d6f0bf258sg-missivenpartition=0.9, seed=42)dh-unibe/image-text_reichsanzeiger-gt @ 9ede9f6c5a3c29e762040b2922ef55355caa2ef1partition=0.9, seed=42)dh-unibe/image-text_gt-fraktur @ eebb83d4e297a9e5b1c86e2ec2b1a2a305d3ad13partition=0.9, seed=42)dh-unibe/image-text_fibeln @ 552ae4dd2669193b614ba3f19c4c13088c87f0b3partition=0.9, seed=42)dh-unibe/image-text_dtgt @ 3d91f02647680fa30cd1a659d1ec03f66b929628partition=0.9, seed=42)dh-unibe/image-text_dach-gt @ ec1faf7d900cbc78609c4ad7db152bbbf8d21951partition=0.9, seed=42)dh-unibe/image-text_weisthuemer @ bb92c7525f93ff1f7c3ae215ef8f8c6eff834922partition=0.9, seed=42)Materialized from that selection: 1,681 pages, 60,743 transcribed lines, 58,101 training samples.
base_model: microsoft/trocr-base-handwritten
epochs: 3
batch_size: 8
accumulate_grad_batches: 1
lrate: 5.0e-05
lr_scheduler: cosine
warmup_ratio: 0.1
weight_decay: 0.0
max_grad_norm: 1.0
optim: adamw_torch
gradient_checkpointing: true
max_new_tokens: 256
beam_size: 4
length_penalty: 1.0
no_repeat_ngram_size: 3
eval_samples: 500
seed: 42
workers: 8
device: cuda:0
wandb_run: null
precision: bf16
| engine | trocr |
| base model | dh-unibe/trocr-kurrent-XVI-XVII |
| training job | 20261026T064540Z-trocr-german-kurrent-xvi-xix-from-kurrent-v2 |
| code | evaluated with 50f7c4fb0cea |
| trained | 2026-10-26T09:54:27.466469+00:00 |
| weights | config.json, generation_config.json, merges.txt, model.safetensors, preprocessor_config.json, special_tokens_map.json, tokenizer.json, tokenizer_config.json, training_args.bin, training_summary.json, vocab.json |
metadata.json in this repo is the record the trainer wrote, verbatim: the full request, the parsed metrics and the job id.
| energy at the cards | 0.443 kWh (443 Wh) |
| CO2e, market-based | 9 g at 20 g/kWh |
| CO2e, location-based | 51 g at 115 g/kWh (Swiss consumer mix) |
Since 2016 the university has procured electricity from 100 % renewable sources (hydro, biomass, solar), predominantly Swiss (source), which is the market-based factor. The location-based figure is what the local grid mix would have cost and is reported beside it, as the GHG Protocol asks when a market instrument is used.
This is a floor. It excludes data-centre overhead (PUE unknown; typical values 1.1–1.6 would raise it by that factor); it counts GPU draw only — CPU, memory, storage and network are not metered; it counts operation only — manufacture of the hardware is not apportioned; it a card was shared with another process, so part of this energy is not this run's.
Weights: config.json (engine: trocr).
This repo: mit — see LICENSE.
Base model dh-unibe/trocr-kurrent-XVI-XVII: mit (per its hub card). These weights are a derivative of it.
dh-unibe/trocr-kurrent-XVI-XVII