Pith. sign in

REVIEW 2 major objections 6 minor 26 references

Transcribing Spanish Texts from the Past: Experiments with Transkribus, Tesseract and Granite

T0 review · 2 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A 2-billion-parameter multimodal model, fine-tuned on a consumer GPU with QLoRA, produced the team's strongest OCR transcriptions in the 2025 PastReader shared task, placing second in five of eight official metrics.

desk verdict A modest, honest shared-task report whose real payload is the competitive Granite 2B fine-tuning result on consumer hardware; the rank claim needs the full leaderboard but the measurements are worth taking seriously. read the letter →

arxiv 2507.04878 v1 pith:K4RL4FTL submitted 2025-07-07 cs.CV cs.CL

classification cs.CVcs.CL
keywords OCRHistoricalDocumentsDigitalHumanitiesMultimodalTranskribusTesseractGranite3.2-visionQLoRAfine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper reports a three-way comparison of OCR approaches for transcribing 19th- and 20th-century Spanish periodicals in the 2025 PastReader shared task: a web-based transcription service, the traditional Tesseract engine, and a compact multimodal model. The paper's main claim is that a 2-billion-parameter vision-language model, Granite3.2-vision:2b, fine-tuned with QLoRA on a single 16 GB consumer GPU, was the team's strongest submission: it placed second in five of eight official evaluation metrics, including word error rate and all ROUGE variants, with margins to the winner that the authors describe as marginal. The authors also report that fine-tuning Tesseract failed to beat its off-the-shelf baseline and that the training dataset contains labelling inconsistencies and distributional differences between splits. Together these results suggest that small multimodal models fine-tuned on ordinary hardware are a practical route for historical OCR, provided the evaluation data is reliable.

What carries the argument

The load-bearing mechanism is the fine-tuned compact vision-language model Granite3.2-vision:2b, a 2-billion-parameter multimodal model whose vision tower processes document images in 384×384 crops. Fine-tuning uses QLoRA, which loads the base model in 4-bit NF4 precision and updates only lightweight LoRA adapters injected into the language model's projection layers, leaving the vision tower and output head in full precision; with a per-device batch size of 1, 8 gradient-accumulation steps, and bfloat16, this fits the whole run into 16 GB of VRAM. The model is prompted through a chat-format dataset with a Spanish-specific OCR system prompt instructing it to return only raw text, and page images are resized to 414×585 with aspect-ratio-preserving padding to stay near the processor's crop size. This mechanism is what carries the paper's claim that a compact multimodal model can rival much larger systems on semantic and word-level metrics while remaining trainable on ordinary hardware.

What would settle it

Recompute the official metrics after re-annotating the test-set ground truth with explicit philological transcription rules—covering diacritics, historical spellings, and line-break decisions—and see whether Granite3.2-vision:2b still places second in WER and the ROUGE variants; if the margins to the winner move beyond the reported thousandths-level gaps, the claimed ranking is an artifact of the noisy ground truth. A complementary check is to measure inter-annotator agreement on a sample of the ground-truth files, since low agreement would directly undermine the metric comparisons.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that Granite3.2-vision:2b—a 2-billion-parameter open multimodal model—can be fine-tuned for end-to-end transcription of historical Spanish documents using QLoRA 4-bit quantization and LoRA adapters on a consumer GPU with 16 GB of VRAM, and the resulting system is 'arguably our strongest submission overall'. In the official leaderboard it took second place in WER and in all four ROUGE variants, with scores within a few thousandths of the winning system, and third in BLEU; only character-level Levenshtein distance and normalised edit distance trailed, placing fourth. The paper attributes the character-level gap to the model's tendency to join hyphen-split words, 'correct' historical Spanish spellings, and hallucinate in illegible areas, and it notes preliminary evidence that dev/train and test splits differ in source diversity and background colour. The same run also demonstrates the feasibility of the whole pipeline—PDF-to-PNG conversion, padding to a 414×585 resolution, prompt-based chat formatting, and quantised fine-tuning—on consumer hardware.

Load-bearing premise

The rankings depend on the shared task's ground-truth transcriptions being consistently annotated, but the authors themselves report finding labelling inconsistencies during fine-tuning, so if those inconsistencies distort the metrics the central comparison could change.

Editorial extensions

If this is right

  • A 2B-parameter multimodal model fine-tuned with QLoRA on a 16 GB consumer GPU can reach second place in five of eight official OCR metrics, so competitive historical transcription no longer requires high-performance computing.
  • Tesseract fine-tuning on this dataset did not improve over the off-the-shelf model, suggesting that traditional OCR transfer learning plateaus unless the training data is clean, consistent, and well aligned.
  • For Transkribus, preserving line breaks helps precision metrics (Levenshtein, WER, NED), while joining lines helps semantic metrics (BLEU, ROUGE), showing that output formatting alone shifts scores across metric families.
  • Granite's best results are in word-level and semantic metrics, so further gains should target literal character accuracy: diacritics, historical spellings, and hyphenated line breaks.
  • The apparent distribution mismatch between dev/train and test sets means future fine-tuning should control for source diversity and background colour before drawing conclusions about model generalisation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural testable extension: once the authors release their planned transcription guidelines, re-running the same QLoRA recipe on a cleaned dataset should specifically improve Levenshtein and NED, since current Granite errors concentrate in diacritics and historical spelling 'corrections'.
  • The paper's line-break findings imply that OCR evaluation campaigns should report metrics on both raw and normalised text, because a model's formatting choices can change BLEU and ROUGE scores without changing transcription content.
  • If the recipe transfers, the same 2B-model-plus-QLoRA pipeline could be applied to other low-resource historical languages; the open question, which the paper does not test, is whether its success depends on Spanish being close enough to the model's English-centric pretraining.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper reports the GRESEL team's participation in the IberLEF 2025 PastReader shared task, which asks systems to transcribe historical Spanish documents from PDF images. Three approaches are compared: Transkribus using the Coloso Español model, Tesseract both off-the-shelf and fine-tuned with tesstrain, and a 2B-parameter Granite3.2-vision model fine-tuned with QLoRA on an RTX 5080. Using the official leaderboard results (Table 3), the authors find that the fine-tuned Granite run places second in five of eight metrics, including WER and all ROUGE variants, and they interpret this as evidence that a compact multimodal model fine-tuned on consumer hardware can be competitive for historical-document OCR. The paper also reports that Tesseract fine-tuning did not improve over the baseline, analyzes failure patterns on difficult pages, discusses the effect of line-break preservation on different metrics, and includes energy/emission estimates. The authors are careful to label several explanations as hypotheses and to acknowledge dataset-quality issues.

Significance. If the ranking claim holds, the paper provides a useful empirical data point: a 2B-parameter vision-language model fine-tuned with QLoRA on a single consumer GPU can outperform traditional OCR engines on WER and semantic overlap metrics for historical Spanish texts, while remaining much cheaper than large-scale HPC training. The paper is also transparent about limitations: it explicitly states that ground-truth inconsistencies exist, that hyperparameter exploration was minimal, that the train/test distribution may differ, and that some observed phenomena are only hypotheses. The practical engineering detail—PDF-to-PNG conversion, padding to fixed resolution, NF4 quantization, gradient checkpointing, and LoRA adapter placement—will be helpful to practitioners who want to reproduce or adapt this pipeline. The main value is therefore a reproducible system-description contribution with an honest assessment of what did and did not work.

major comments (2)
  1. [Section 4.1, Table 3] The claim that GRESEL2_run1 achieved "a solid second place in five out of the eight evaluation metrics" is only verifiable if Table 3 is the complete official leaderboard for the task. The table lists seven rows, but the manuscript never states that this is the set of all submitted runs, and the cited overview paper [1] is not reproduced. If any other team's run lies between OCRTITS_run1 and GRESEL2_run1 on WER or any ROUGE metric, the stated rank would be incorrect. Please add an explicit sentence that Table 3 contains the complete official leaderboard (with the total number of participating teams/runs), or otherwise qualify the rank claim.
  2. [Sections 2.2, 3.2, and 4.1] The authors themselves report "inconsistencies in labelling" in the training data (Section 3.2) and state that "it is essential that the dataset is clean and well organised, as well as the application of systematic transcription rules" (Section 2.2), which they say is still pending. Since all leaderboard metrics are computed against this ground truth, the central quantitative claims—including the Granite second-place result—are potentially affected by label noise or inconsistent transcription conventions. The manuscript should explicitly discuss how such inconsistencies could bias the comparison, for example by favoring models that reproduce the annotators' errors or by penalizing models that transcribe the page literally. This is a load-bearing caveat rather than a stylistic one, and it should appear in or near Section 4.1.
minor comments (6)
  1. [Section 4.1] The text says "even achieving second place in BLUE"; the metric name should be BLEU. The same typo appears in the description of the Transkribus joined-lines run.
  2. [Section 3.3] The 90/10 random split of the combined training and development sets is described as producing "training" and "testing" partitions. Since the official test set is used later, calling this second split "testing" is confusing; recommend "validation" or "held-out subset."
  3. [Section 3.2] The conclusion that fine-tuned Tesseract "consistently" underperforms the baseline is mainly supported by Figure 3 on the development set, but the figure's caption does not identify which metrics are shown or how scores are normalized. Adding axis labels and a legend would make the comparison easier to verify.
  4. [Section 4.1, Figure 8] The emission analysis is described as a "simulated 10-hour execution window," but the method for estimating energy consumption and CO2 is not specified. A one-sentence description of the estimation formula or tool would improve reproducibility.
  5. [Section 3.3, Table 1] The hyperparameter row "Dataset text field \"\"" is unclear; it likely refers to the text column being empty for image-only inputs. Please clarify what this setting does.
  6. [Throughout] There are several minor wording and formatting errors, including "an useful point," "As it can bee seen," "theColoso Español" (missing space), and the email address in the author affiliation block appears garbled. These do not affect the technical content but should be corrected in a final revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Granite result is anchored to an external official leaderboard, not derived from the paper's own fitted inputs.

full rationale

The paper's central empirical claim is the ranking of GRESEL2_run1 on the PastReader 2025 test set, as reported in Table 3 with results provided by the task organization. The fine-tuned Granite model was trained on a random 90/10 split of the development and training sets and then run on the test set released later by the organization (Section 3.3); the authors explicitly state that hyperparameters remained largely unexplored and were mostly defaults or initial guesses, so no parameter was fitted to the test labels. The leaderboard is an external benchmark, and the paper does not reconstruct the ranking from any fitted quantity of its own. The self-citations to the companion papers [3] and [4] are descriptive references to separate system descriptions rather than load-bearing evidence for the validity of the current results. The dataset citation [21] includes two of the present authors as contributors, but the dataset and official evaluation are products of the shared task organization, and no argument in the paper reduces to that citation. The observation that Table 3 lists only seven runs and therefore cannot by itself prove 'second place' against the complete official leaderboard is an evidence-completeness concern, not circularity. Likewise, the authors' admission of labeling inconsistencies in the ground truth is a data-quality limitation that could distort all metric comparisons, but it does not make any claimed prediction equivalent to its input by construction. No equation or claimed result in the paper reduces to a fitted parameter renamed as a prediction, and no uniqueness or validity claim is imported from the authors' prior work.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

Central claim is an empirical benchmark comparison, so the ledger records the modeling and data assumptions rather than fitted parameters. Hyperparameters are listed as free parameters because they were chosen by hand and affect the outcome. No new entities are introduced.

free parameters (6)
  • Granite learning rate = 1e-4
    Chosen by default/minimal testing (Section 3.3, Table 1), affects fine-tuning outcome.
  • Number of training epochs = 1
    Default value; no exploration of longer training.
  • Gradient accumulation steps = 8
    Set to fit in 16GB VRAM; affects effective batch size.
  • Target image resolution = 414x585
    Reduced from original to fit memory; may cause information loss (Section 3.3).
  • Train/dev split ratio for fine-tuning = 90/10 random
    Random split of combined train+dev sets; seed not reported.
  • Tesseract page segmentation mode = psm 6
    Chosen based on 'empirical recommendations' (Section 3.2).
assumptions (4)
  • domain assumption The official evaluation metrics (WER, BLEU, ROUGE family, edit distances) are valid proxies for OCR transcription quality.
    Comparison of runs relies entirely on these metrics (Section 4.1).
  • domain assumption Ground-truth transcriptions in the PastReader dataset are correct and consistently transcribed.
    Authors themselves flag potential labelling inconsistencies and missing transcription rules (Sections 2.2 and 3.2).
  • domain assumption The shared task test set is representative of historical Spanish documents and is not leaked into fine-tuning.
    Authors observe a possible distribution mismatch between dev/train and test sets (Section 4.2).
  • domain assumption QLoRA 4-bit quantization with NF4 preserves enough model quality for OCR.
    Quantization was required to fit in 16GB VRAM; no ablation against full precision (Section 3.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transcribing Spanish Texts from the Past: Experiments with Transkribus, Tesseract and Granite." pith.science (2026). https://pith.science/paper/K4RL4FTL

@misc{pith2026250704878,
  author       = {Pith},
  title        = {Pith review of: Transcribing Spanish Texts from the Past: Experiments with Transkribus, Tesseract and Granite},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K4RL4FTL}},
  note         = {Machine review of arXiv:2507.04878}
}
read the original abstract

This article presents the experiments and results obtained by the GRESEL team in the IberLEF 2025 shared task PastReader: Transcribing Texts from the Past. Three types of experiments were conducted with the dual aim of participating in the task and enabling comparisons across different approaches. These included the use of a web-based OCR service, a traditional OCR engine, and a compact multimodal model. All experiments were run on consumer-grade hardware, which, despite lacking high-performance computing capacity, provided sufficient storage and stability. The results, while satisfactory, leave room for further improvement. Future work will focus on exploring new techniques and ideas using the Spanish-language dataset provided by the shared task, in collaboration with Biblioteca Nacional de Espa\~na (BNE).

Figures

Figures reproduced from arXiv: 2507.04878 by the authors.

Figure 1
Figure 1. Examples of pages with poor performance. based on coherent choices from a philological and palaeographic point of view. For this reason, we are preparing annotation and transcription guidelines based on the issues detected in the documents of this training dataset for future work. 3. Experiments Our three main experiments are described in the following lines. The first one, Transkribus, was used for inference using … view at source ↗
Figure 2
Figure 2. Comparison of the working directory structure before and after applying fine-tuning. The fine-tuning process began with transcription alignment, where all corrected text transcriptions were renamed to adhere to the .gt.txt naming convention required by Tesseract for supervised learning. This step ensures that each transcription is correctly paired with its corresponding image during training. Next, image conversion … view at source ↗
Figure 3
Figure 3. Tesseract charts comparing performance between fine-tuned and non-fine-tuned model using the development set. was observed that certain files appeared in multiple “top 10” worst-performing lists. Since this analysis was performed manually, attention was focused on the files that occurred in more than four of these lists. Because these files consistently showed poor performance across several metrics, they are assume… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Images corresponding to the worst-performing files common to both the baseline and the fine-tuned model. Granite3.2-vision:2b was more prone to wrongly interpret or misspell Spanish words, probably because its training data is in English language. We would like to give…
Figure 5
Figure 5. Figure 5: Dataset structure required by Granite3.2-vision:2b. As mentioned, it also requires a system prompt. It is more detailed and contains precise instructions of what is expected from the model. The system prompt used can be seen in [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Prompt for OCR with Spanish documents. It is worth noting that a challenging aspect was the use of RAM memory for loading the dataset image files as pixel values. The employed gaming pc features 32gb of RAM memory. Despite the fact that this is not a low amount of RAM …
Figure 7
Figure 7. Figure 7: Scattered plots for each metric sorted by best scores [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Energy and emission estimations: Granite3.2-vision:2b and Tesseract. 4.2. Qualitative insights Upon inspecting the output generated by Granite model, it was quite satisfactory overall, but some patterns were identified. For instance, when a word is split across lines b…
Figure 9
Figure 9. Figure 9: File 2466 carrera: y tu' c^i^isitícás 4^^^^ ^altí) sas estén terafeen^^^^^míl^^ ^ _^,^^jtres ro-a|Yi3|Cido....^ QhX eso pq, j(íof ,vÍda^T^;:;p^4^ ^mc^,b»j,flu^^^^^ jBQJ ,9,acex,^,hX-ifi madíe ¡tü^ir.o |i;ln ,p^- -í^f'íiwsifl^óofo^lie na¿^í(ÍQ gjp'9,na9,cp ¿9S./Tiay(^^^…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 16 canonical work pages

  1. [1]

    Montejo-Ráez, E

    A. Montejo-Ráez, E. Sánchez-Nogales, G. Expósito-Álvarez, L. A. Ureña-López, M. T. Martín- Valdivia, J. Collado-Montañez, I. Cabrera-de Castro, M. V. Cantero-Romero, R. Ortuño-Casanova, Overview of pastreader shared task in iberlef 2025: Transcribing texts from the past, Procesamiento del Lenguaje Natural 75 (2025)

  2. [2]

    J. Á. González-Barba, L. Chiruzzo, S. M. Jiménez-Zafra, Overview of IberLEF 2025: Natural Language Processing Challenges for Spanish and other Iberian Languages, in: Proceedings of the Iberian Languages Evaluation Forum (IberLEF 2025), co-located with the 41st Conference of the Spanish Society for Natural Language Processing (SEPLN 2025), CEUR-WS. org, 2025

  3. [3]

    J. Macicior-Mitxelena, Transcribing History with Tesseract: A Monomodal OCR Approach in the PastReader 2025 Shared Task, in: Proceedings of the Iberian Languages Evaluation Forum (IberLEF 2025) co-located with the 41st Conference of the Spanish Society for Natural Language Processing (SEPLN 2025), volume TBD of CEUR Workshop Proceedings, CEUR-WS, Zaragoza...

  4. [4]

    Y. A. Torterolo-Orta, M. Miguez-Lamanuzzi, Fine-Tuning a Compact Multimodal Model on Consumer-Grade Hardware at PastReader 2025, in: Proceedings of the Iberian Languages Evalua- tion Forum (IberLEF 2025) co-located with the 41st Conference of the Spanish Society for Natural Language Processing (SEPLN 2025), volume TBD of CEUR Workshop Proceedings , CEUR-W...

  5. [5]

    Sánchez-Salido, A

    E. Sánchez-Salido, A. Menta, A. García-Serrano, Seeking information in spanish historical newspa- pers: The case of diario de madrid (18th and 19th centuries)., DHQ: Digital Humanities Quarterly (2023)

  6. [6]

    J. J. Lastra-Díaz, J. Goikoetxea, M. A. Hadj Taieb, A. Garcia-Serrano, M. Ben Aouicha, E. Agirre, D. Sánchez, A large reproducible benchmark of ontology-based methods and word embeddings for word similarity, Information Systems 96 (2021) 101636. URL: https://www.sciencedirect.com/ science/article/pii/S0306437920301058. doi:https://doi.org/10.1016/j.is.2020.101636

  7. [7]

    Garcia-Arias, A

    E. Garcia-Arias, A. Garcia-Serrano, Creación de un modelo de descripciones de imágenes especial- izado en arqueología griega (pendiente de publicación), Procesamiento del Lenguaje Natural 75 (2025)

  8. [8]

    Garcia-Serrano, X

    A. Garcia-Serrano, X. Benavent, R. Granados, J. M. Goñi-Menoyo, Some results using different approaches to merge visual and text-based features in clef’08 photo collection, in: C. Peters, T. Dese- laers, N. Ferro, J. Gonzalo, G. J. F. Jones, M. Kurimo, T. Mandl, A. Peñas, V. Petras (Eds.), Evaluating Systems for Multilingual and Multimodal Information Acc...

Show all 26 references
  1. [9]

    J. L. Martínez-Fernández, J. V. Román, A. M. Garcia-Serrano, J. C. González-Cristóbal, Combining textual and visual features for image retrieval, in: C. Peters, F. C. Gey, J. Gonzalo, H. Müller, G. J. F. Jones, M. Kluck, B. Magnini, M. de Rijke (Eds.), Accessing Multilingual I...

  2. [10]

    Moreno-Sandoval, J

    A. Moreno-Sandoval, J. Porta, B. Carbajo-Coronado, Y. Torterolo, D. Samy, The financial document causality detection shared task (FinCausal 2025), in: C.-C. Chen, A. Moreno-Sandoval, J. Huang, Q. Xie, S. Ananiadou, H.-H. Chen (Eds.), Proceedings of the Joint Workshop of the 9t...

  3. [11]

    Porta-Zamorano, Y

    J. Porta-Zamorano, Y. Torterolo-Orta, A. Moreno-Sandoval, LLI-UAM Team at FinancES 2023: Noise, Data Augmentation and Hallucinations, in: Proceedings of the Iberian Languages Evaluation Forum (IberLEF 2023) co-located with the Conference of the Spanish Society for Natural Lang...

  4. [12]

    Y. A. Torterolo-Orta, S. M. Roseti, A. Moreno-Sandoval, Assessing a literary rag system with a human-evaluated synthetic qa dataset generated by an llm: Experiments with knowledge graphs, Procesamiento del Lenguaje Natural 75 (2025). In press

  5. [13]

    H. Liu, C. Li, Q. Wu, Y. J. Lee, Visual instruction tuning, 2023. URL: https://arxiv.org/abs/2304.08485. arXiv:2304.08485

  6. [14]

    Grattafiori, A

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, A. Yang, A. Fan, A. Goyal, A. Hartshorn, A. Yang, A. Mitra, A. Sravankumar, A. Korenev, A. Hinsvark, (...), Z. Ma, The llama 3 herd of models, 2024. URL: http...

  7. [15]

    G. Team, A. Kamath, J. Ferret, S. Pathak, N. Vieillard, R. Merhej, S. Perrin, T. Matejovicova, A. Ramé, M. Rivière, L. Rouillard, T. Mesnard, G. Cideron, J. bastien Grill, S. Ramos, E. Yvinec, M. Casbon, E. Pot, I. Penchev, (...), L. Hussenot, Gemma 3 technical report, 2025. U...

  8. [16]

    Poznanski, J

    J. Poznanski, J. Borchardt, J. Dunkelberger, R. Huff, D. Lin, A. Rangapur, C. Wilhelm, K. Lo, L. Soldaini, olmocr: Unlocking trillions of tokens in pdfs with vision language models, 2025. URL: https://arxiv.org/abs/2502.18443. arXiv:2502.18443

  9. [17]

    P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, Y. Fan, K. Dang, M. Du, X. Ren, R. Men, D. Liu, C. Zhou, J. Zhou, J. Lin, Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution, arXiv preprint arXiv:2409.12191 (2024)

  10. [18]

    G. V. Team, L. Karlinsky, A. Arbelle, A. Daniels, A. Nassar, A. Alfassi, B. Wu, E. Schwartz, D. Joshi, J. Kondic, N. Shabtay, P. Li, R. Herzig, S. Abedin, S. Perek, S. Harary, U. Barzelay, A. R. Goldfarb, A. Oliva, B. Wieles, (...), R. Feris, Granite vision: a lightweight, ope...

  11. [19]

    Hegghammer, Ocr with tesseract, amazon textract, and google document ai: a benchmarking experiment, Journal of Computational Social Science 5 (2022) 861–882

    T. Hegghammer, Ocr with tesseract, amazon textract, and google document ai: a benchmarking experiment, Journal of Computational Social Science 5 (2022) 861–882

  12. [20]

    Fleischhacker, W

    D. Fleischhacker, W. Goederle, R. Kern, Improving ocr quality in 19th century historical documents using a combined machine learning based approach, arXiv preprint arXiv:2401.07787 (2024)

  13. [21]

    Montejo-Ráez, E

    A. Montejo-Ráez, E. Sánchez Nogales, G. Expósito Álvarez, A. Ureña López, M. T. Martín-Valdivia, J. Collado-Montañez, I. Cabrera de Castro, M. V. Cantero Romero, A. García Serrano, R. Or- tuño Casanova, Y. A. Torterolo Orta, Pastreader 2025, https://doi.org/10.5281/zenodo.15084265,

  14. [22]

    Schwartz, Fine-tuning granite vision with trl and peft (lora), https://colab.research.google

    E. Schwartz, Fine-tuning granite vision with trl and peft (lora), https://colab.research.google. com/github/huggingface/cookbook/blob/main/notebooks/en/fine_tuning_granite_vision_sft_trl. ipynb, 2024. URL: https://colab.research.google.com/github/huggingface/cookbook/blob/main...

  15. [23]

    Zhang, O

    J. Zhang, O. Liu, T. Yu, J. Hu, W. Neiswanger, Euclid: Supercharging multimodal llms with synthetic high-fidelity visual descriptions, arXiv preprint arXiv:2412.08737 (2024)

  16. [24]

    Dettmers, A

    T. Dettmers, A. Pagnoni, A. Holtzman, L. Zettlemoyer, Qlora: Efficient finetuning of quantized llms, 2023. URL: https://arxiv.org/abs/2305.14314. arXiv:2305.14314

  17. [25]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, Lora: Low-rank adapta- tion of large language models, 2021. URL: https://arxiv.org/abs/2106.09685. arXiv:2106.09685

  18. [26]

    Narbona, S

    A. Narbona, S. Ros, Beyond traditional ocr: Exploring the efficiency of llms in document processing, in: Proceedings of the Iberian Languages Evaluation Forum (IberLEF 2025) co-located with the 41st Conference of the Spanish Society for Natural Language Processing (SEPLN 2025)...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.