RVLCdip Document Images with Text for Classification

A dataset containing document images and associated text, structured with 10 examples per class for classification tasks. The dataset was created by author 'nielsr' and was last updated on Hugging Face in August 2022. It is derived from the RVL-CDIP collection, focusing on a subset with balanced examples.

Use Cases

• Train a document classifier using the image modality to predict the document class label.
• Fine-tune a vision-language model by aligning image features with extracted text features from the documents.
• Benchmark multimodal fusion techniques that combine image and text data for improved classification accuracy.
• Use the 10-examples-per-class structure for few-shot learning experiments in document analysis.

Strengths

• Contains a balanced subset with 10 examples per class, useful for controlled experiments.
• Provides multimodal data (image and text) for document analysis tasks.
• Derived from the established RVL-CDIP benchmark dataset for document classification.

Limitations

• Small sample size per class (10 examples) limits training of large models from scratch.
• The temporal coverage and geographic origin of the source documents are unknown.
• Potential class imbalance or bias from the original RVL-CDIP source is not addressed.

Provenance

Source

Derived from the RVL-CDIP (Ryerson Vision Lab Complex Document Information Processing) dataset.

Collection Method

Subsampled to create a balanced set with 10 examples per document class.

Freshness

Last updated in August 2022; static snapshot.