indic-deva-ocr-eval
收藏资源简介:
Indic Deva Eval 是一个专为评估印度天城文(Devanagari)光学字符识别(OCR)系统性能而构建的综合性基准数据集。该数据集旨在广泛覆盖多种OCR应用场景,包括印刷文档页面、独立数字、裁剪出的单词图像以及手写文本,以全面测试模型在不同条件下的识别能力。数据集包含多个子集的混合样本,具体构成如下:25%来自`devanagari_page_ocr`的测试集,25%来自`indic_vision_bench_deva_ocr`的OCR测试集,30%来自`indic_mozhi_deva_word_ocr`(其中印地语和马拉地语各占15%的测试集),15%来自`hindi_handwritten_word_ocr`的测试集,以及5%来自`devanagari_digits_mixed`的训练集。这种混合确保了数据在来源和任务类型上的多样性。数据以两种主要格式提供:1)核心的`.ocr.jsonl`文件,其中包含`image`(图像文件相对路径)、`ocr`(真实文本标签)、`source_repo`(来源仓库)以及语言/来源信息等列;2)可选的`.sharegpt.json`文件,适用于模型微调和评估,包含`messages`和`images`字段。每个样本都有一个唯一的`id`标识符。数据集支持尼泊尔语(ne)、印地语(hi)和马拉地语(mr)三种使用天城文书写的语言。它主要适用于图像到文本的OCR任务,可作为模型评估基准或微调数据源。需要注意的是,数据集由`scripts/sample_ocr_eval_sets.py`脚本从GLM微调工作区生成,在使用时,若原始数据源仅提供训练分割,应确保将确定的保留行ID排除在监督微调(SFT)或训练运行之外。
Indic Deva Eval is a comprehensive benchmark dataset specifically constructed to evaluate the performance of Devanagari optical character recognition (OCR) systems. This dataset aims to broadly cover diverse OCR application scenarios, including printed document pages, isolated numerals, cropped word images, and handwritten text, to comprehensively test the recognition capabilities of models under varying conditions. The dataset comprises mixed samples from multiple subsets, with the specific composition as follows: 25% from the test split of `devanagari_page_ocr`, 25% from the OCR test split of `indic_vision_bench_deva_ocr`, 30% from `indic_mozhi_deva_word_ocr` (with Hindi and Marathi each accounting for 15% of its test split), 15% from the test split of `hindi_handwritten_word_ocr`, and 5% from the training split of `devanagari_digits_mixed`. This mixture ensures diversity in both data sources and task types. The data is provided in two main formats: 1) The core `.ocr.jsonl` file, which contains columns such as `image` (relative path to the image file), `ocr` (ground-truth text label), `source_repo` (source repository), and language/source information; 2) The optional `.sharegpt.json` file, suitable for model fine-tuning and evaluation, which includes the `messages` and `images` fields. Each sample has a unique `id` identifier. The dataset supports three Devanagari-written languages: Nepali (ne), Hindi (hi), and Marathi (mr). It is primarily intended for image-to-text OCR tasks, serving as a model evaluation benchmark or fine-tuning data source. It should be noted that the dataset was generated from the GLM fine-tuning workspace via the `scripts/sample_ocr_eval_sets.py` script. When using it, if the original data source only provides the training split, it is necessary to ensure that the confirmed reserved line IDs are excluded from supervised fine-tuning (SFT) or training runs.
数据集概述:Indic Deva Eval
- 数据集名称:Indic Deva Eval
- 任务类别:图像到文本(image-to-text)
- 语言:尼泊尔语(ne)、印地语(hi)、马拉地语(mr)
- 标签:OCR、天城文(Devanagari)、GLM-OCR
- 漂亮名称:Indic Deva Eval
数据集描述
该数据集是一个广泛的天城文OCR基准测试,涵盖印刷页面、数字、单词裁剪和手写内容。
关键信息
- 主要原始文件:
*.ocr.jsonl,包含图像、OCR文本、来源仓库、语言和来源列。 - 可选微调/评估文件:
*.sharegpt.json,包含消息和图像。
核心列
- id:唯一样本标识符
- image:图像文件的相对路径
- ocr:真实文本标签
数据来源组合
| 来源 | 拆分 | 占比 |
|---|---|---|
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% |
配置
- 配置名称:
default - 数据文件:
- 拆分:
test - 路径:
indic_deva_eval.viewer.ocr.parquet
- 拆分:
备注
- 该数据集由GLM微调工作区中的
scripts/sample_ocr_eval_sets.py生成。 - 如果某个来源仅提供了训练拆分,请确保将确定的保留行ID排除在SFT/训练运行之外。




