sora 13274243a0 Bump vendored EvalScope and add K3-ready DPV4 configs.
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>
2026-09-02 07:30:48 +00:00

3.6 KiB

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_antonym subset downloads a small public MP4 archive for quick validation
  • Full benchmark evaluation can be requested by setting subset_list to additional MVBench subsets
  • Time-bounded records keep start/end metadata and add a short segment instruction to the prompt

Properties

Property Value
Benchmark Name mvbench
Dataset ID PKU-Alignment/MVBench
Paper Paper
Tags MCQ, MultiModal, Video
Metrics accuracy
Default Shots 0-shot
Evaluation Split train

Data Statistics

Metric Value
Total Samples 4,000

Per-Subset Statistics:

Subset Samples Prompt Mean Prompt Min Prompt Max
action_antonym 200 N/A N/A N/A
action_count 200 N/A N/A N/A
action_localization 200 N/A N/A N/A
action_prediction 200 N/A N/A N/A
action_sequence 200 N/A N/A N/A
character_order 200 N/A N/A N/A
counterfactual_inference 200 N/A N/A N/A
egocentric_navigation 200 N/A N/A N/A
episodic_reasoning 200 N/A N/A N/A
fine_grained_action 200 N/A N/A N/A
fine_grained_pose 200 N/A N/A N/A
moving_attribute 200 N/A N/A N/A
moving_count 200 N/A N/A N/A
moving_direction 200 N/A N/A N/A
object_existence 200 N/A N/A N/A
object_interaction 200 N/A N/A N/A
object_shuffle 200 N/A N/A N/A
scene_transition 200 N/A N/A N/A
state_change 200 N/A N/A N/A
unexpected_action 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}

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
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)