Keep K3 suite selection and report-schema scoring in bash, merge K3/vision dataset_args into dpv4 yamls, and pin EvalScope at 735d920ee911 with local patches. Co-authored-by: Cursor <cursoragent@cursor.com>
4.1 KiB
4.1 KiB
MeasureBench
Overview
MeasureBench is a comprehensive benchmark for evaluating the ability of vision-language models (VLMs) to read values from measuring instruments. It covers both real-world photographs and synthetically generated images of 26 instrument types across 4 design categories.
Task Description
- Task Type: Free-form Visual Question Answering (instrument reading)
- Input: An image of a measuring instrument + a reading question
- Output: The instrument's current reading (numeric value or time, with unit)
- Domains: Ammeters, clocks, thermometers, scales, speedometers, and 21 more instrument types
Key Features
- 2,442 total samples across two splits: real_world (1,272) and synthetic_test (1,170)
- 26 instrument types, 4 design categories (dial, digital, analog, linear)
- Accepts a tolerance interval around the correct value rather than requiring an exact match
- For clocks: handles both 12-hour and 24-hour ambiguity via multiple valid intervals
- Unit recognition is evaluated separately from numeric accuracy
Evaluation Notes
- Default splits: real_world and synthetic_test (treated as separate subsets)
- Primary metric: Accuracy (
accuracy) —all_correct: number and unit both correct - Secondary metrics: number_acc (numeric only), unit_acc (unit only)
- Two evaluators:
interval_matching(single valid range) andmulti_interval_matching(e.g. clock AM/PM) - Model output is expected in the format
Answer: <value> <unit>on the last line image_typeis recorded in each sample's metadata; per-type results are visible in thesubset_keycolumn of review files but are not separately selectable viasubset_list- Paper | GitHub
Properties
| Property | Value |
|---|---|
| Benchmark Name | measure_bench |
| Dataset ID | evalscope/MeasureBench |
| Paper | Paper |
| Tags | MultiModal, QA, Reasoning |
| Metrics | accuracy, number_acc, unit_acc |
| Default Shots | 0-shot |
| Evaluation Split | real_world |
Data Statistics
| Metric | Value |
|---|---|
| Total Samples | 2,442 |
| Prompt Length (Mean) | 150.9 chars |
| Prompt Length (Min/Max) | 126 / 215 chars |
Per-Subset Statistics:
| Subset | Samples | Prompt Mean | Prompt Min | Prompt Max |
|---|---|---|---|---|
real_world |
1,272 | 153.83 | 131 | 215 |
synthetic_test |
1,170 | 147.71 | 126 | 192 |
Image Statistics:
| Metric | Value |
|---|---|
| Total Images | 2,442 |
| Images per Sample | min: 1, max: 1, mean: 1 |
| Resolution Range | 108x79 - 3025x1599 |
| Formats | jpeg, png |
Sample Example
Subset: real_world
{
"input": [
{
"id": "1341f508",
"content": [
{
"image": "[BASE64_IMAGE: jpeg, ~75.8KB]"
},
{
"text": "What is the reading of the instrument?\nProvide your final answer on the last line in the format: Answer: <value> <unit>. For example: Answer: 42.5 A"
}
]
}
],
"target": "",
"id": 0,
"group_id": 0,
"subset_key": "ammeter",
"metadata": {
"question_id": "ammeter_0",
"image_type": "ammeter",
"design": "dial",
"evaluator": "interval_matching",
"evaluator_kwargs": "{\"interval\": [9.5, 9.7], \"units\": [\"A\", \"Ampere\"]}"
}
}
Prompt Template
No prompt template defined.
Usage
Using CLI
evalscope eval \
--model YOUR_MODEL \
--api-url OPENAI_API_COMPAT_URL \
--api-key EMPTY_TOKEN \
--datasets measure_bench \
--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=['measure_bench'],
limit=10, # Remove this line for formal evaluation
)
run_task(task_cfg=task_cfg)