MVBench#
Overview#
MVBench is a public multimodal video understanding benchmark covering temporal perception,
attribute/state reasoning, symbolic ordering, and high-level cognition. This native adapter uses
the ModelScope PKU-Alignment/MVBench mirror by default, which provides JSON annotations plus
optimized video archives.
Task Description#
Task Type: Video multiple-choice question answering
Input: Video + question + answer choices
Output: Single correct answer letter
Subsets: 20 MVBench tasks; the default smoke-test subset is
action_antonym
Evaluation Notes#
Default configuration uses 0-shot evaluation
Primary metric: Accuracy
The default
action_antonymsubset downloads a small public MP4 archive for quick validationFull benchmark evaluation can be requested by setting
subset_listto additional MVBench subsetsTime-bounded records keep start/end metadata and add a short segment instruction to the prompt
Properties#
Property |
Value |
|---|---|
Benchmark Name |
|
Dataset ID |
|
Paper |
|
Tags |
|
Metrics |
|
Default Shots |
0-shot |
Evaluation Split |
|
Data Statistics#
Metric |
Value |
|---|---|
Total Samples |
4,000 |
Per-Subset Statistics:
Subset |
Samples |
Prompt Mean |
Prompt Min |
Prompt Max |
|---|---|---|---|---|
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
|
200 |
N/A |
N/A |
N/A |
Sample Example#
Sample example not available.
Prompt Template#
Prompt Template:
Answer the following multiple choice question. The last line of your response should be of the following format: 'ANSWER: [LETTER]' (without quotes) where [LETTER] is one of {letters}. Think step by step before answering.
{question}
{choices}
Extra Parameters#
Parameter |
Type |
Default |
Description |
|---|---|---|---|
|
|
|
Dataset repository ID or local dataset root for MVBench annotations and videos. |
|
|
|
Dataset hub used to load annotations and video archives. Choices: [‘huggingface’, ‘modelscope’, ‘local’] |
|
|
`` |
Optional dataset revision; leave empty to use the hub default. |
Usage#
Using CLI#
evalscope eval \
--model YOUR_MODEL \
--api-url OPENAI_API_COMPAT_URL \
--api-key EMPTY_TOKEN \
--datasets mvbench \
--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=['mvbench'],
dataset_args={
'mvbench': {
# subset_list: ['action_antonym', 'action_count', 'action_localization'] # optional, evaluate specific subsets
# extra_params: {} # uses default extra parameters
}
},
limit=10, # Remove this line for formal evaluation
)
run_task(task_cfg=task_cfg)