---
dataset_info:
  features:
  - name: instruction
    dtype: string
  - name: annotation_id
    dtype: string
  - name: actions
    sequence:
    - name: action_id
      dtype: string
    - name: obs_before_path
      dtype: string
    - name: obs_after_path
      dtype: string
    - name: obs_before
      dtype: image
    - name: obs_after
      dtype: image
    - name: operation
      dtype: string
    - name: bbox
      struct:
      - name: x
        dtype: float32
      - name: 'y'
        dtype: float32
      - name: width
        dtype: float32
      - name: height
        dtype: float32
    - name: metadata
      struct:
      - name: repr
        dtype: string
      - name: text
        dtype: string
  - name: source
    dtype: string
  - name: platform
    dtype: string
  - name: action_space
    sequence: string
  - name: is_success
    dtype: bool
  splits:
  - name: train
    num_bytes: 2343867236
    num_examples: 345
  download_size: 1425594267
  dataset_size: 2343867236
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: mit
task_categories:
- visual-question-answering
language:
- en
---

# IDM-Multiple

IDM-Multiple is part of the IDMBench in AgentStudio for evaluating agent's ability to learn from Internet-scale videos without action labels. Specifically, this dataset requires VLMs to answer an action sequence within a video clip of screen recoding. Please note that this dataset is a test set rather than a training set. Therefore, please do not use it for training. More details are provided in the [project page](https://ltzheng.github.io/agent-studio).