Indic Deva OCR Eval: A Benchmark for Devanagari Script Recognition

A benchmark for evaluating Optical Character Recognition (OCR) systems on Devanagari script. The dataset is hosted on Hugging Face by the author himalaya-ai and was last updated on June 2, 2026. It likely contains a mix of printed pages, digits, word crops, and handwriting samples.

Use Cases

• Benchmarking OCR model accuracy on printed Devanagari pages.
• Evaluating digit recognition performance within Indic documents.
• Testing OCR robustness on isolated word crops.
• Assessing handwriting recognition capabilities for Devanagari script.

Strengths

• Covers multiple challenging OCR sub-tasks: printed pages, digits, word crops, and handwriting.
• Source mix is specified, with devanagari_page_ocr and indic_vision_bench_deva_ocr each contributing 25% to a default/test split.
• Provides structured data files: a main *.ocr.jsonl file and an optional fine-tuning/eval *.sharegpt.json file.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

huggingface datasets platform, author himalaya-ai.

Collection Method

Aggregated from multiple source repositories including devanagari_page_ocr, indic_vision_bench_deva_ocr, and indic_mozhi_deva_word_ocr.

Freshness

Last updated 2026-06-02 15:54:55; freshness should be verified.