nougat-base

Converts scanned or digital images of scientific papers, technical documents, and academic PDFs into structured Markdown text using a vision-encoder-decoder architecture. The model employs a Swin Transformer vision encoder to extract spatial features from document images, then decodes them into LaTeX-compatible Markdown using a transformer decoder trained on arXiv papers. This enables preservation of mathematical equations, tables, and hierarchical document structure in machine-readable format.

Enables efficient batch processing of multiple document images through the Hugging Face Transformers library's pipeline abstraction, supporting dynamic batching and automatic device placement (CPU/GPU). The model integrates with the standard transformers.pipeline() interface, allowing developers to load the model once and process multiple images with automatic tensor batching, memory management, and optional GPU acceleration without manual CUDA code.

Extracts text from scientific document images while preserving mathematical equations in LaTeX format, using a decoder trained on arXiv papers where equations are annotated with their source LaTeX. The model learns to recognize equation regions in images and generate corresponding LaTeX code rather than attempting to OCR equations as plain text, enabling downstream tools to render or parse equations correctly.

Implements a modular vision-encoder-decoder architecture where a Swin Transformer encoder extracts hierarchical visual features from document images, and a transformer decoder generates Markdown text token-by-token. The encoder processes images at multiple scales (4×, 8×, 16×, 32×) to capture both fine details and document structure, while the decoder uses cross-attention to align generated text with visual features, enabling structured output generation.

Loads model weights from Hugging Face Hub using the safetensors format, which provides secure deserialization without arbitrary code execution risks. The model is distributed as safetensors files instead of pickle, preventing malicious code injection during model loading. Integration with transformers library enables automatic format detection and loading without explicit format specification.

Integrates with Hugging Face Hub for automatic model discovery, downloading, and caching. The model is hosted on Hub with versioning support, allowing developers to specify model revisions and automatically cache downloaded weights locally. Integration with transformers library enables one-line model loading with automatic Hub authentication, version management, and cache directory configuration.

Trained on arXiv papers spanning multiple languages and scientific domains, enabling the model to handle documents in English, Chinese, Japanese, and other languages common in academic publishing. The decoder learns language-specific tokenization and formatting conventions through exposure to diverse arXiv papers, supporting multilingual Markdown output with proper character encoding.