---
configs:
- config_name: default
  data_files:
  - split: test
    path: indic_deva_eval.viewer.ocr.parquet
task_categories:
- image-to-text
language:
- ne
- hi
- mr
tags:
- ocr
- devanagari
- glm-ocr
pretty_name: Indic Deva Eval
---

# indic_deva_eval

Broad Indic Devanagari OCR benchmark across printed pages, digits, word crops, and handwriting.

- Repo: `himalaya-ai/indic-deva-ocr-eval`
- Task: `indic_devanagari_ocr`
- Main raw file: `*.ocr.jsonl` with `image`, `ocr`, `source_repo`, and language/provenance columns.
- Optional fine-tuning/eval file: `*.sharegpt.json` with `messages` and `images`.

## Core Columns

- `id`: unique sample identifier
- `image`: relative path to the image file
- `ocr`: ground-truth text label

## Source Mix

- `devanagari_page_ocr` `default/test`: 25%
- `indic_vision_bench_deva_ocr` `ocr/test`: 25%
- `indic_mozhi_deva_word_ocr` `hindi/test`: 15%
- `indic_mozhi_deva_word_ocr` `marathi/test`: 15%
- `hindi_handwritten_word_ocr` `default/test`: 15%
- `devanagari_digits_mixed` `default/train`: 5%

## Notes

Generated by `scripts/sample_ocr_eval_sets.py` from the GLM fine-tuning workspace.
If a source only exposes a train split, keep the deterministic held-out row ids out of SFT/training runs.