OCRBench-v2

Overview

OCRBench v2 is a large-scale bilingual text-centric benchmark with the most comprehensive set of OCR tasks (4x more than OCRBench v1), covering 31 diverse scenarios including street scenes, receipts, formulas, diagrams, and more.

Task Description

Task Type: Optical Character Recognition and Document Understanding
Input: Image + OCR/document question
Output: Text recognition, extraction, or analysis result
Languages: English and Chinese (bilingual)

Key Features

10,000 human-verified question-answering pairs
31 diverse scenarios (street scene, receipt, formula, diagram, etc.)
High proportion of difficult samples
Comprehensive OCR task coverage
Bilingual (English and Chinese) evaluation

Evaluation Notes

Default configuration uses 0-shot evaluation
Evaluates on test split
Requires: apted, distance, Levenshtein, lxml, Polygon3, zss packages
Simple accuracy metric

Properties

Property	Value
Benchmark Name	`ocr_bench_v2`
Dataset ID	evalscope/OCRBench_v2
Paper	N/A
Tags	`Knowledge`, `MultiModal`, `QA`
Metrics	`acc`
Default Shots	0-shot
Evaluation Split	`test`

Data Statistics

Metric	Value
Total Samples	10,000
Prompt Length (Mean)	155.62 chars
Prompt Length (Min/Max)	6 / 1863 chars

Per-Subset Statistics:

Subset	Samples	Prompt Mean	Prompt Min	Prompt Max
`APP agent en`	300	36.9	16	84
`ASCII art classification en`	200	174.62	162	193
`key information extraction cn`	400	32.45	9	162
`key information extraction en`	400	506.23	327	1261
`key information mapping en`	300	664.02	429	1647
`VQA with position en`	300	557.69	539	612
`chart parsing en`	400	67	67	67
`cognition VQA cn`	200	15.91	6	48
`cognition VQA en`	800	44.73	14	179
`diagram QA en`	300	118.23	54	356
`document classification en`	200	314	314	314
`document parsing cn`	300	43	43	43
`document parsing en`	400	51	51	51
`formula recognition cn`	200	19	19	19
`formula recognition en`	400	58.39	53	60
`handwritten answer extraction cn`	200	39.75	22	60
`math QA en`	300	119	119	119
`full-page OCR cn`	200	31	31	31
`full-page OCR en`	200	91	91	91
`reasoning VQA en`	600	80.45	26	256
`reasoning VQA cn`	400	163.96	62	633
`fine-grained text recognition en`	200	155.63	152	156
`science QA en`	300	387.54	159	1863
`table parsing cn`	300	139.4	79	211
`table parsing en`	400	65.39	57	134
`text counting en`	200	113.87	101	127
`text grounding en`	200	361.5	357	379
`text recognition en`	800	37.26	29	104
`text spotting en`	200	446	446	446
`text translation cn`	400	230.88	96	291

Image Statistics:

Metric	Value
Total Images	10,000
Images per Sample	min: 1, max: 1, mean: 1
Resolution Range	19x10 - 3912x21253
Formats	jpeg

Sample Example

Subset: APP agent en

{
  "input": [
    {
      "id": "51292e32",
      "content": [
        {
          "text": "What is the wrong answer 2?"
        },
        {
          "image": "[BASE64_IMAGE: jpeg, ~121.7KB]"
        }
      ]
    }
  ],
  "target": "[\"enabled\", \"on\"]",
  "id": 0,
  "group_id": 0,
  "subset_key": "APP agent en",
  "metadata": {
    "question": "What is the wrong answer 2?",
    "answers": [
      "enabled",
      "on"
    ],
    "eval": "None",
    "dataset_name": "rico",
    "type": "APP agent en",
    "bbox": null,
    "bbox_list": null,
    "content": null
  }
}

Prompt Template

Prompt Template:

{question}

Usage

Using CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets ocr_bench_v2 \
    --limit 10  # Remove this line for formal evaluation

Using Python

from evalscope import run_task
from evalscope.config import TaskConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['ocr_bench_v2'],
    dataset_args={
        'ocr_bench_v2': {
            # subset_list: ['APP agent en', 'ASCII art classification en', 'key information extraction cn']  # optional, evaluate specific subsets
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)

4.7 KiB Raw Blame History