Source-linked AI summary
Typhoon OCR: Open Vision-Language Model For Thai Document Extraction
Surapon Nonesung, Natapong Nitarach, Teetouch Jaknamon, Pittawat Taveekitworachai, Kunat Pipatanakul
TL;DR
Thai document understanding remains challenging because complex script properties, irregular layouts, and limited multimodal training data constrain existing VLMs. Typhoon OCR fine-tunes open vision-language backbones with task-aligned real and synthetic data for Thai and English extraction. Its models perform competitively with proprietary systems, while V1.5 improves average performance over V1 despite using 2B rather than 7B parameters.
Problem
Thai document extraction is limited by complex script properties, irregular layouts, and scarce image-aligned structured annotations compared with high-resource languages.
Method
Typhoon OCR fine-tunes open vision-language backbones on a task-aligned corpus combining curated real documents and synthetic data, with separate Default and Structure supervision modes.
Results
Typhoon OCR V1.5 achieves higher average performance than V1 despite using 2B rather than 7B parameters, while Typhoon OCR performs competitively with proprietary models across Thai document categories.
Takeaways & Limitations
Open, compact, task-adapted VLMs can provide strong Thai document extraction and layout reconstruction with reduced computational overhead.
Takeaways & Limitations
Figure understanding remains a limitation, with Thai Books showing lower performance across models because illustrations and non-standard figures increase layout ambiguity.
Abstract
from arXiv · showhide
Document extraction is a core component of digital workflows, yet existing vision-language models (VLMs) predominantly favor high-resource languages. Thai presents additional challenges due to script complexity from non-latin letters, the absence of explicit word boundaries, and the prevalence of highly unstructured real-world documents, limiting the effectiveness of current open-source models. This paper presents Typhoon OCR, an open VLM for document extraction tailored for Thai and English. The model is fine-tuned from vision-language backbones using a Thai-focused training dataset. The dataset is developed using a multi-stage data construction pipeline that combines traditional OCR, VLM-based restructuring, and curated synthetic data. Typhoon OCR is a unified framework capable of text transcription, layout reconstruction, and document-level structural consistency. The latest iteration of our model, Typhoon OCR V1.5, is a compact and inference-efficient model designed to reduce reliance on metadata and simplify deployment. Comprehensive evaluations across diverse Thai document categories, including financial reports, government forms, books, infographics, and handwritten documents, show that Typhoon OCR achieves performance comparable to or exceeding larger frontier proprietary models, despite substantially lower computational cost. The results demonstrate that open vision-language OCR models can achieve accurate text extraction and layout reconstruction for Thai documents, reaching performance comparable to proprietary systems while remaining lightweight and deployable.
1 Introduction
Thai document understanding is difficult because the language combines complex script properties with diverse, irregular document layouts, while existing VLMs and multimodal datasets remain oriented toward high-resource languages. Typhoon OCR addresses this gap as an open VLM tailored to Thai document extraction.
- Thai’s stacked diacritics, context-dependent vowels, and absent word boundaries hinder reliable text segmentation and recognition.
- Dense, irregular layouts in forms, financial records, receipts, and tabular reports complicate accurate extraction and structural reconstruction.
- General-purpose VLMs frequently produce recognition errors, layout misinterpretations, and semantic inconsistencies on Thai documents.
- Thai lacks large-scale multimodal datasets pairing document images with structured textual and semantic annotations.
- Typhoon OCR is an open VLM for end-to-end Thai and English document understanding, including text extraction, layout reconstruction, and document-level semantic modeling.
- The work proposes a Thai-focused data curation and synthesis pipeline and evaluates the models against open and proprietary baselines.
2 Typhoon OCR
Typhoon OCR combines task-aligned supervision, structure-sensitive operating modes, and a multi-stage corpus construction pipeline to extract and reconstruct Thai and English documents. Its evaluations show strong performance on structured documents and competitive results with substantially smaller models, while Thai Books remain challenging because of visual elements.
- 2 Typhoon OCR: Typhoon OCR supports end-to-end Thai and English extraction with text recognition and layout-aware content reconstruction.
- Dataset Creation Pipeline: Default Mode and Structure Mode provide different output granularity for loosely structured versus highly structured documents.
- Dataset Creation Pipeline: The Structure Mode corpus pipeline extracts text with OCR, restructures it using prompted VLMs, applies consistency checks, and verifies a subset manually.
- Dataset Creation Pipeline: The multi-stage pipeline produces scalable, consistent, high-quality supervision for fine-tuning under both modes.
- Dataset Statistics: 77,029 document samples span Thai and English infographics, government forms, financial reports, books, handwritten materials, scans, and other documents.
- Results and Discussion: Typhoon OCR consistently outperforms GPT-4o and Gemini 2.5 Flash on financial reports and government forms, with gains most pronounced when PDF metadata are available.
- Results and Discussion: Thai Books have lower performance across models because illustrations and non-standard figures increase ambiguity in figure representation and layout interpretation.
- Results and Discussion: The 3B variant achieves results comparable to the 7B model on several tasks, particularly government forms, supporting constrained deployment settings.
3 Typhoon OCR V1.5
Typhoon OCR V1.5 simplifies data construction and deployment while improving document extraction across diverse Thai and English document scenarios. Its compact 2B model outperforms the 7B V1 on average, though proprietary models remain stronger for visually heterogeneous documents.
- Model and Data Construction: Thai-translated VQA preserves general vision-language grounding, while synthetic documents address scarce annotations for complex layouts, equations, charts, rare vocabulary, and typographic variation.Synthetic generation combines Thai vocabulary, visual elements, mathematical expressions, and document-level image augmentation.
- Training Corpus: The training corpus combines retained Typhoon OCR data, VQA, DocLayNet supervision, and synthetic documents, with 37.6% allocated to synthetic data.The retained V1 corpus contributes 53.7%, VQA 2.2%, and DocLayNet 6.4%.
- Training Setup: The model uses full-parameter supervised fine-tuning of Qwen3-VL 2B with long-context multimodal support and quantization-aware training for efficient low-precision inference.Training uses a 16,384-token maximum sequence length and selects the final checkpoint by validation performance.
- Results: Across Thai document categories, V1.5 achieves higher average performance than V1 despite using 2B rather than 7B parameters.The evaluation compares both Typhoon versions with two frontier VLM baselines across multiple categories.
- Results: V1.5 performs strongest on structured government forms and financial reports, but proprietary models retain lower character-level error rates on infographics and handwritten forms.The structured-document gains include higher BLEU and ROUGE-L and lower Levenshtein distances.
4 Conclusion
Typhoon OCR improves Thai document understanding across transcription, layout reconstruction, and structural consistency, with V1.5 matching or exceeding larger proprietary systems at lower inference cost. Limitations remain for severely degraded inputs, broader language coverage, higher-level reasoning, and academic benchmark assessment.
- Typhoon OCR improves transcription accuracy, layout reconstruction, and structural consistency across document-understanding tasks.
- Typhoon OCR V1.5 uses 2B parameters while matching or exceeding larger proprietary models across multiple document categories.This reduces inference cost relative to larger proprietary systems.
- Performance degrades on severely degraded inputs, including low-resolution images, motion blur, and occlusions.The paper suggests improved data recipes or explicit modeling of noise and capture artifacts as future remedies.
- The model series currently supports primarily Thai and English and focuses on document extraction rather than higher-level reasoning tasks.Future work is proposed for other low-resource languages, diagram understanding, and structured information extraction.
- Broader assessment on academic benchmarks such as ThaiOCRBench is identified as a next step for understanding model capabilities.