FC-AMF-OCR: 9.3 Million OCR-Processed French Financial Document Images

A collection of 9.3 million document images processed through Optical Character Recognition (OCR) using the docTR library. The dataset is derived from the AMF-PDF dataset, part of the Finance Commons collection, and was created by lightonai. The dataset card was last updated on September 23, 2024.

Use Cases

• Train or fine-tune OCR models based on a large-scale collection of financial document images.
• Benchmark document layout analysis algorithms using images derived from regulatory filings.
• Develop financial text extraction pipelines using the OCR-processed text output.
• Pre-train multimodal models for financial document understanding on image-text pairs.

Strengths

• Contains 9.3 million images, indicating a large-scale resource.
• Processed with a specific OCR library (docTR), providing a consistent text extraction method.
• Derived from a known source collection (Finance Commons AMF-PDF), suggesting a defined domain.

Limitations

• The description notes that native text annotations in a related dataset suffer from imperfections, which may imply OCR quality issues.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the text data is unknown, which may limit suitability assessment.

Provenance

Source

Derived from the AMF-PDF dataset within the Finance Commons collection.

Collection Method

Images processed through Optical Character Recognition (OCR) using the docTR library.

Freshness

Last updated 2024-09-23 16:30:39; freshness should be verified.

Geography

Likely France, based on the AMF (Autorité des Marchés Financiers) source, but not explicitly confirmed.