Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Finetuning Vision-Language Models as OCR Systems for Low-Resource Languages: A Case Study of Manchu

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Fine-tuning LLaMA-3.2-11B on synthetic Manchu word images produces an OCR system that reaches 93.1% word accuracy on real handwritten 18th-century documents, far above a CRNN baseline's 72.5%.

desk verdict Plausible VLM fine-tuning recipe for Manchu OCR, but the headline 93.1% real-world accuracy is not credible because the checkpoint was picked on the same 218-image test set used to report it. read the letter →

arxiv 2507.06761 v1 pith:TZQMRINH submitted 2025-07-09 cs.CV

classification cs.CV
keywords ManchuOCRvision-languagemodellow-resourcelanguagesyntheticdatahandwrittentextrecognitionparameter-efficientfine-tuningdigitalhumanities
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 sets out to show that open vision-language models, fine-tuned only on synthetic rendered images, can become accurate optical character recognition systems for Manchu, a critically endangered script with large historical archives and no robust public OCR. It reports that LLaMA-3.2-11B, fine-tuned on 60,000 synthetic word images with parameter-efficient training, reaches 98.3% word accuracy on synthetic test images and keeps 93.1% on 218 handwritten words from an 18th-century Qing document. By contrast, a conventional CRNN baseline scores 99.8% on synthetic data but collapses to 72.5% on the same handwritten test set, while smaller Qwen2.5-VL models fall to 43.1% and 23.9%. A sympathetic reader should care because the result suggests that low-cost, locally runnable synthetic-data training can open up large Manchu-language archives to historians and linguists.

What carries the argument

The load-bearing mechanism is parameter-efficient LoRA fine-tuning of a pretrained vision-language model on a large synthetic dataset, combined with a structured dual-output prompt that forces the model to emit both the Manchu script and its romanized transliteration. That dual output matters because the romanization channel gives the model a stable phonetic target that can disambiguate visually similar Manchu glyphs, especially vowels such as A, E, and I that are nearly identical in isolation and only distinguishable by context and diacritics. The paper attributes the LLaMA model's robust real-world performance to this contextual, linguistically grounded decoding, which the CRNN's purely visual sequence model lacks.

What would settle it

Run the released model on a held-out set of handwritten Manchu words from several different documents, scribes, and centuries, with manual segmentation and ground truth produced by a Manchu specialist, without preprocessing the images to resemble synthetic data. If word accuracy on that unmodified sample falls substantially below 93.1%, the paper's synthetic-to-real transfer claim fails for real archive conditions.

Watch

Extended reading notes

Core claim

The central discovery claim is that the pretrained linguistic representations in a large vision-language model provide the inductive bias needed for synthetic-to-real transfer in low-resource script OCR. Fine-tuned on synthetic Manchu word images with a fixed instruction that asks for both the Manchu script and its romanized transliteration, the LLaMA-3.2-11B model achieves a character error rate of 0.0024 and 98.3% word accuracy on synthetic validation data, and 0.0219 CER with 93.1% word accuracy on real handwritten samples from Neige Cangben Manwen Laodang. The same fine-tuning protocol applied to Qwen2.5-VL-3B and Qwen2.5-VL-7B produces strong synthetic performance but severe degradation on real handwriting, and a CRNN baseline optimized for the synthetic distribution also fails to transfer. The paper presents this as evidence that, for endangered languages with limited real annotated data, fine-tuned open VLMs can outperform both task-specialized classical architectures and commercial closed models, which achieved 0% word accuracy in the paper's limited comparison.

Load-bearing premise

The load-bearing premise is that the 218 real handwritten word images used for testing, automatically segmented from one 18th-century document and then preprocessed to approximate the synthetic training style, are a fair, correctly labeled sample of real Manchu historical handwriting; if the segmentation, preprocessing, or labels distort the domain, the reported 93.1% accuracy will not transfer to actual archives.

Editorial extensions

If this is right

  • Historians and linguists could run accurate Manchu OCR on a single workstation or accessible cloud environment, removing the need for specialized infrastructure.
  • Synthetic data alone, without large real annotated corpora, can apparently bridge to real handwritten documents for at least one endangered script, pointing to a reusable recipe for other low-resource languages.
  • Model scale and pretraining quality appear decisive: the 11B LLaMA model transfers far better than the 3B and 7B Qwen models under identical fine-tuning, suggesting that OCR for low-resource scripts should favor larger open VLMs.
  • Because the model also outputs romanized transliteration, the OCR pipeline can directly support full-text search, indexing, and downstream computational analysis of Manchu archives.
  • The open release of code and weights makes the system reproducible and extendable, which is essential for endangered-language preservation work that depends on long-term community access.

Reading between the lines

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

  • Beyond the paper, the 93.1% real-world figure is likely optimistic as a general estimate: it rests on only 218 handwritten word images from a single document, automatically segmented and preprocessed to resemble the synthetic training data, so true archive-level accuracy on varied scribes and centuries is untested.
  • The dual-output prompt hints at a self-consistency signal that the paper does not exploit at inference time: decoding could be improved by generating both script and romanization and enforcing agreement between them, a testable extension for future work.
  • The method's transferability to other endangered scripts may depend on Manchu's alphabetic syllabary with relatively regular glyph forms; cursive scripts, line-level recognition, or heavily degraded manuscripts may require additional real data or different prompts.
  • A direct next experiment is to evaluate the released model on full pages or multi-word lines rather than pre-segmented words, since real archive workflows need page-level OCR and current segmentation errors would compound recognition errors.
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

5 major / 6 minor

Summary. The paper proposes fine-tuning three open-source vision-language models (LLaMA-3.2-11B, Qwen2.5-VL-7B, Qwen2.5-VL-3B) on 60,000 synthetic Manchu word images using LoRA-based parameter-efficient training, and evaluates them on a synthetic validation set and a 218-image real handwritten test set from the Neige Cangben Manwen Laodang. The central claims are that LLaMA-3.2-11B reaches 98.3% word accuracy and 0.0024 CER on synthetic data while retaining 93.1% word accuracy on real handwritten words, and that this substantially outperforms a CRNN baseline, which drops from 99.8% synthetic accuracy to 72.5% on the real test set. The paper frames the approach as a transferable, cost-effective OCR framework for endangered languages and releases code and model weights.

Significance. If the reported accuracies were valid, the work would be a practically valuable demonstration that open-weight VLMs can be adapted to low-resource scripts with modest compute, and it would provide a useful baseline for digital-humanities OCR. The manuscript also has concrete strengths: it releases code and model weights, uses a parameter-efficient LoRA setup that is replicable on a single GPU, and addresses a genuine gap in Manchu historical-document processing. However, the headline real-world number is produced by selecting the final checkpoint on the same 218-image test set that is then used to report performance, which makes the reported 93.1% an upwardly biased extreme value rather than an unbiased estimate. Because the central comparative and cross-domain claims rest on this protocol, the current evidence does not support them as stated.

major comments (5)
  1. [Section 5.1] The selection of the final model checkpoints is performed on the same 218-image real-world test set that is then used to report the headline results. The text states that 'we evaluated all checkpoints on our real-world test dataset and selected those achieving the highest word accuracy performance,' and for LLaMA-3.2-11B the selected checkpoint is at step 21,000. With binary per-word outcomes and n=218, the maximum accuracy over roughly 20 checkpoints is an order statistic that is systematically higher than the true accuracy, so 93.1% is not an unbiased estimate of cross-domain generalization. The authors must select checkpoints on a validation set disjoint from the final test set, or otherwise correct for selection, and the same requirement applies to the CRNN comparison if it is used to support the comparative claim.
  2. [Section 3] The construction of the synthetic evaluation sets is not specified tightly enough to rule out contamination. The paper first selects 60,000 images from a larger dataset for training, then states that 15,000 synthetic images are used as a validation set during training and that 1,000 synthetic images are randomly sampled for final assessment, but it never states whether the validation and final-assessment images are disjoint from the 60,000 training images or how they relate to the original 750,000/25,000 split. If all three sets are random draws from the same pool, repeated word identities or near-duplicate rendered images can appear in both training and evaluation, inflating the synthetic accuracies in Table 2. Please specify the sampling procedure and verify pairwise disjointness at the image and word-token level.
  3. [Section 3] The real-world test set does not measure the claimed domain gap as cleanly as stated. The 218 handwritten words are automatically segmented from a single source document (Neige Cangben Manwen Laodang), and the images are then preprocessed to 'approximate the visual style and format of the synthetic training data.' This preprocessing deliberately reduces the very distribution shift that the paper claims to evaluate, and no assessment of segmentation accuracy is reported. The authors should either evaluate on unmodified real images, report the effect of each preprocessing step on accuracy, and provide a test set drawn from multiple documents, or explicitly reframe the claim as performance on preprocessed crops from a single manuscript rather than on 'real-world historical documents.'
  4. [Section 5.3 and Appendix B] The CRNN comparison is not protocol-matched. The VLM checkpoints were selected on the real test set, whereas the CRNN checkpoint was selected on the synthetic validation set (epoch 48), so the two test-set numbers are not directly comparable. Appendix B also describes the CRNN training data only as '60,000 validation samples,' which is ambiguous about whether it is the same preprocessed set used for the VLMs. If the real test images were converted to the VLM synthetic style while the CRNN was trained on a different image representation, the CRNN is additionally disadvantaged. Please align the data splits, preprocessing, and checkpoint-selection rule for both families of models before drawing conclusions about VLM superiority.
  5. [Section 5.2] All headline numbers are point estimates on a single 218-sample test set with no confidence intervals or statistical comparison. For a binomial proportion, 203/218 = 93.1% has a standard error of roughly 1.7 percentage points, but because the reported value is a maximum selected from multiple checkpoints, ordinary interval formulas do not apply. The paper should report confidence intervals and, ideally, evaluate on additional independent real-image sets so that the comparison with CRNN (72.5%) and the robustness claim are statistically grounded.
minor comments (6)
  1. [Section 3] The text says 'As shown in Figures 2' but refers to a single figure; the citation should be 'Figure 2'.
  2. [Section 5.2] The table title calls the 1,000-sample set 'Validation Dataset', while Section 3 states that 15,000 synthetic images are used as the validation set during training; the terminology should be made consistent.
  3. [Section 5.4] The phrase 'closed-domain model (GPT-4.1)' is imprecise; the intended meaning is a closed-source or API-only model, not a model restricted to a closed domain.
  4. [Section 4.4] The metrics section should state explicitly that CER, word accuracy, and F1 are computed on the Manchu-script output only and that the Roman transliteration is excluded from the reported numbers.
  5. [Figure 6] The x-axis is labeled 'Validation Test' without clearly distinguishing which tick corresponds to which condition; the figure should use separate axis labels for Validation and Test and a legend that identifies the two models.
  6. [Appendix B] Table 7 lists 'Dataset 60,000 validation samples' for the CRNN; this should say 'training samples' or otherwise clarify the exact training split and preprocessing.

Circularity Check

1 steps flagged · score 6.0 of 10

The 93.1% real-world accuracy is a maximum over checkpoints selected on the same 218-image test set, so the headline result is partly determined by the test data rather than independently predicted.

  1. fitted input called prediction [Section 5.1 (Model Selection and Training Stability); Section 5.2 (Overall Performance)]
    "Following training completion, we evaluated all checkpoints on our real-world test dataset and selected those achieving the highest word accuracy performance. ... LLaMA-3.2-11B maintained robust performance with 93.1% word accuracy and 0.983 F1 score, representing only a modest 5.2 percentage point accuracy decline from validation."

    The central reported real-world accuracy is not an independent estimate: the same 218 handwritten samples are used both to choose the best checkpoint (by saving checkpoints every 1,000 steps and selecting the highest word accuracy on that test set) and then to report the final number. The 93.1% figure is therefore the maximum over the evaluated checkpoints on the very samples used for the headline, not a held-out prediction. With n=218 per-sample binary outcomes, this test-set-based model selection inflates the reported accuracy relative to true generalization. The selected checkpoint is a fitted input that is then presented as the model's real-world performance, so the headline claim partially reduces to the selection criterion.

full rationale

No equation-level derivation is present, and there are no load-bearing self-citations: the synthetic dataset, base models, and prior Manchu OCR work are external sources. The only significant circularity is in the evaluation loop. The paper explicitly evaluates checkpoints on the real-world test set and selects the one with highest word accuracy, then reports that same test set's accuracy (93.1%) as the main result. This makes the central claim an optimized fit to the test labels rather than an unbiased prediction of cross-domain transfer. The additional preprocessing of the real test images to 'approximate the visual style and format of the synthetic training data' is a real validity caveat, but it is better understood as a domain-gap narrowing than as construction-level circularity; it does not by itself make the reported number equivalent to an input. Overall, the paper's key comparative numbers are partially inputs to the selection procedure, giving a mild-to-moderate circularity burden rather than a full derivation-level circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or linguistic entities. The central claim rests on the source synthetic dataset, the automatic segmentation of a single historical document, assumed-correct test labels, and LoRA as a reliable adaptation method.

free parameters (6)
  • LoRA rank = 64
    Set to 64 for all models, chosen by preliminary experiments reported in Section 4.3.1.
  • LoRA alpha = 64
    Set equal to rank with no ablation, from Section 4.3.1.
  • Learning rate = 1e-4
    Chosen by preliminary experiments and applied uniformly, from Section 4.3.1.
  • Number of training epochs = 15 (3B), 10 (7B), 5 (11B)
    Hand-selected schedule based on model size and assumed sample efficiency, from Section 4.3.1.
  • Training sample count = 60,000
    Random subset of the 750k synthetic corpus, chosen by compute budget; no study of the effect of this number is provided, from Section 3.
  • Checkpoint selection criterion = best real-test word accuracy
    Selection rule uses the real test set, making reported test performance an optimized value rather than an unbiased estimate, from Section 5.1.
assumptions (5)
  • domain assumption The synthetic Manchu dataset by Zhang Zhuohui correctly renders the Manchu script and its romanizations.
    The paper relies on this dataset for all training and synthetic evaluation; any errors propagate to reported performance, from Section 3.
  • domain assumption The OpenCV-based segmentation algorithm extracts complete and correctly ordered Manchu words from the scanned document.
    Real test samples are generated by this algorithm; incorrect segmentation would corrupt images or labels, from Section 3.
  • domain assumption The ground-truth labels for the 218 handwritten test words are correct.
    No human verification of the extracted words or labels is described in Section 3.
  • domain assumption LoRA fine-tuning preserves and builds on the base VLMs' pretrained visual and linguistic capabilities.
    The transfer advantage over CRNN is attributed to pretrained representations; if LoRA disrupts them, the interpretation changes, from Section 4.3.1.
  • standard math Levenshtein distance is an appropriate character error metric for Manchu script strings.
    Treats each Unicode character as a token; a standard OCR metric, from Section 4.4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Finetuning Vision-Language Models as OCR Systems for Low-Resource Languages: A Case Study of Manchu." pith.science (2026). https://pith.science/paper/TZQMRINH

@misc{pith2026250706761,
  author       = {Pith},
  title        = {Pith review of: Finetuning Vision-Language Models as OCR Systems for Low-Resource Languages: A Case Study of Manchu},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TZQMRINH}},
  note         = {Machine review of arXiv:2507.06761}
}
read the original abstract

Manchu, a critically endangered language essential for understanding early modern Eastern Eurasian history, lacks effective OCR systems that can handle real-world historical documents. This study develops high-performing OCR systems by fine-tuning three open-source vision-language models (LLaMA-3.2-11B, Qwen2.5-VL-7B, Qwen2.5-VL-3B) on 60,000 synthetic Manchu word images using parameter-efficient training. LLaMA-3.2-11B achieved exceptional performance with 98.3\% word accuracy and 0.0024 character error rate on synthetic data, while crucially maintaining 93.1\% accuracy on real-world handwritten documents. Comparative evaluation reveals substantial advantages over traditional approaches: while a CRNN baseline achieved 99.8\% synthetic accuracy, it suffered severe degradation to 72.5\% on real documents. Our approach demonstrates effective synthetic-to-real domain transfer, providing a cost-effective solution deployable on accessible infrastructure. This work establishes a transferable framework for endangered language OCR that removes technical and financial barriers in digital humanities, enabling historians and linguists to process historical archives without specialized computing resources. Code and model weights are available at https://github.com/mic7ch1/ManchuAI-OCR.

Figures

Figures reproduced from arXiv: 2507.06761 by the authors.

Figure 1
Figure 1. manju gisun, meaning ‘Manchu language’. The Manchu script is generally regarded as an alphabetic syllabary, derived from the vertically written Mongolian script, which in turn comes from Old Uyghur. It consists of individual letters representing both vowels and con￾sonants. These letters vary in shape depending on their position within a syllable—initial, medial, or final—and are visually connected in flowing vertic… view at source ↗
Figure 2
Figure 2. Original and preprocessed Manchu word images [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Original hand-written Manchu word from Neige [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Hand-written Manchu word after processing for eval [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Training dynamics and checkpoint selection [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Performance comparison between CRNN and LLaMA-3.2-11B models [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Top misrecognized characters in Manchu predictions [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Multi-Expert Routing for Multi-Domain Low-Resource OCR: A Manchu Case Study

    cs.CV 2026-07 conditional novelty 4.0 of 10

    A style classifier routing Manchu pages to checkpoints from a fine-tuning history matches the per-domain best CERs (0.30/1.57/4.83%), but the specialists were selected on the same frozen test sets.

Reference graph

Works this paper leans on

22 extracted references · 19 canonical work pages · cited by 1 Pith paper

  1. [1]

    Pamela Kyle Crossley and Evelyn S. Rawski. A profile of the manchu language in ch’ing history. Harvard Journal of Asiatic Studies, 53(1):63–102, 1993

  2. [2]

    Evelyn S. Rawski. Presidential address: Reenvi- sioning the qing: The significance of the qing pe- riod in chinese history. The Journal of Asian Stud- ies, 55(4):829–850, 1996. doi: 10.2307/2646525

  3. [3]

    An offline recognition method of handwritten primitive manchu characters based on strokes

    Guang-yuan Zhang, Jing-jiao Li, Rong-wei He, and Ai-xia Wang. An offline recognition method of handwritten primitive manchu characters based on strokes. In Ninth International Workshop on Fron- tiers in Handwriting Recognition, pages 432–437. IEEE, 2004

  4. [4]

    Design and implementation of off line hand- written document recognition system of manchu manuscript

    Ji ZHAO, Jing LI, Guang ZHANG, Jie WAN, et al. Design and implementation of off line hand- written document recognition system of manchu manuscript. Pattern Recognition and Artificial In- telligence, 19(6):801–805, 2006

  5. [5]

    A new recognition method for the handwritten manchu character unit

    Guang-yuan Zhang, Jing-jiao Li, and Ai-xia Wang. A new recognition method for the handwritten manchu character unit. In 2006 International Con- ference on Machine Learning and Cybernetics , pages 3339–3344. IEEE, 2006

  6. [6]

    Synthetic data and dag-svm classifier for segmentation-free manchu word recognition

    Di Huang, Min Li, Ruirui Zheng, Shuang Xu, and Jiajing Bi. Synthetic data and dag-svm classifier for segmentation-free manchu word recognition. In 2017 International Conference on Computing In- telligence and Information System (CIIS) , pages 46–50. IEEE, 2017

  7. [7]

    Manchu word recognition based on convolutional neural network with spatial pyramid pooling

    Min Li, Ruirui Zheng, Shuang Xu, Yu Fu, and Di Huang. Manchu word recognition based on convolutional neural network with spatial pyramid pooling. In 2018 11th International Congress on Image and Signal Processing, BioMedical Engi- neering and Informatics (CISP-BMEI), pages 1–6. IEEE, 2018

  8. [8]

    Segmentation-free multi-font printed manchu word recognition using deep convolutional features and data augmentation

    Ruirui Zheng, Min Li, Jianjun He, Jiajing Bi, and Baochun Wu. Segmentation-free multi-font printed manchu word recognition using deep convolutional features and data augmentation. In 2018 11th Inter- national Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP- BMEI), pages 1–6. IEEE, 2018

Show all 22 references
  1. [9]

    Ocr with the deep cnn model for ligature 10 script-based languages like manchu

    Diandian Zhang, Yan Liu, Zhuowei Wang, and De- pei Wang. Ocr with the deep cnn model for ligature 10 script-based languages like manchu. Scientific pro- gramming, 2021(1):5520338, 2021

  2. [10]

    ManchuOCR: A OCR system for the Manchu script

    tyotakuki. ManchuOCR: A OCR system for the Manchu script. https://github.com/ tyotakuki/ManchuOCR, 2022. Accessed: 2025- 05-29

  3. [11]

    Learn more manchu words with a new visual-language framework.ACM Trans- actions on Asian and Low-Resource Language In- formation Processing, 2024

    Zhiwei Wang, Siyang Lu, Xiang Wei, Run Su, Yingjun Qi, and Wei Lu. Learn more manchu words with a new visual-language framework.ACM Trans- actions on Asian and Low-Resource Language In- formation Processing, 2024

  4. [12]

    Qwen for jawi v1: a model for jawi ocr, 2024

    Miguel Escobar Varela. Qwen for jawi v1: a model for jawi ocr, 2024. URL https://huggingface. co/mevsg/qwen-for-Jawi-v1 . Model created at National University of Singapore

  5. [13]

    Manchu: A textbook for reading documents

    Gertraude Roth Li. Manchu: A textbook for reading documents. University of Hawaii Press, 2000

  6. [14]

    Neige cangben Manwen lao- dang (Old documents in Manchu)

    Editorial Committee of Neige cangben Man- wen laodang, editor. Neige cangben Manwen lao- dang (Old documents in Manchu). Liaoning minzu chubanshe, Shenyang, 2009. Edited by Wu Yuan- feng and Zhang Yu, and others

  7. [15]

    Large ocr model: An empirical study of scaling law for ocr

    Xuewang Ding, Lei Meng, Zhuohan Li, Yan Zhu, Zhentao Xie, Yipeng Zhang, Xiaoyong Wei, and Junchi Yan. Large ocr model: An empirical study of scaling law for ocr. arXiv preprint arXiv:2401.00028, 2024

  8. [16]

    Xmodel-vlm: A simple baseline for multimodal vision language model

    Wanting Xu, Yang Liu, Langping He, Xucheng Huang, and Ling Jiang. Xmodel-vlm: A simple baseline for multimodal vision language model. arXiv preprint arXiv:2405.09215, 2024

  9. [17]

    Qwen2.5-VL: Vision-language model

    Qwen Team and Alibaba Cloud. Qwen2.5-VL: Vision-language model. https://huggingface. co/Qwen/Qwen2.5-VL, 2025

  10. [18]

    The llama 3 herd of models

    Meta AI. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

  11. [19]

    Un- sloth: Finetune LLMs 2x faster with 70% less memory, 2023

    Daniel Han, Michael Han, and Unsloth team. Un- sloth: Finetune LLMs 2x faster with 70% less memory, 2023. URL https://github.com/ unslothai/unsloth. Open-source library for ef- ficient fine-tuning of large language models

  12. [20]

    Tinyllama: An open-source small language model

    Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024

  13. [21]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. URL https: //openreview.net/forum?id=nZeVKeeFYf9

  14. [22]

    Levenshtein

    Vladimir I. Levenshtein. Binary codes capable of correcting deletions, insertions, and reversals. So- viet Physics Doklady, 10(8):707–710, 1966. Trans- lated from Doklady Akademii Nauk SSSR, vol. 163, no. 4, pp. 845–848, 1965. Introduces the Leven- shtein distance algorithm. 1...

Pith tools

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