Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
1.92k
1.92k
label
class label
76 classes
1002
1002
1002
1002
1002
1002
1002
1002
2003
2003
2003
2003
2003
2003
3004
3004
3004
3004
3004
3004
3004
3004
3004
3004
3004
3004
3004
3004
3004
3004
6007
6007
6007
6007
6007
6007
6007
6007
6007
6007
6007
6007
6007
6007
6007
6007
7008
7008
7008
7008
7008
7008
7008
7008
7008
7008
7008
7008
7008
7008
8009
8009
8009
8009
8009
8009
8009
8009
8009
8009
8009
8009
8009
8009
9010
9010
9010
9010
9010
9010
9010
9010
10011
10011
10011
10011
10011
10011
10011
10011
10011
10011
10011
10011
10011
10011
10011
10011
11012
11012
End of preview. Expand in Data Studio

OSWorld-Persona

OSWorld-Persona is a multimodal benchmark for evaluating whether a proactive desktop agent can adapt its decision to offer help to different users. It is derived from OSWorld v2 computer-use trajectories and contains four synthetic user personas with different preferences about when and how an assistant should intervene. The synthetic users are simulated by Kimi-K2.6.

Each decision point contains ordered screenshots, elapsed timestamps, a summarized observation, the user's immediate intent, a binary want_support label, a rationale, and any context required to provide useful support. The full persona used to synthesize each user's labels is available in that persona's metadata.json as ground_truth_persona.

The data synthesis pipeline is open-sourced at NVLabs/deep-personalization.

Intended Use

The dataset supports evaluation of personalized proactive agent. The training split represents prior interaction history available for personalization. The test split measures decisions on later examples for the same persona. This setup evaluates adaptation to known preference profiles; it does not evaluate generalization to entirely unseen users.

Dataset Summary

Persona Train Train yes Test Test yes
automation_first 521 202 239 100
eager_handson 509 390 237 192
learning_oriented 660 37 285 10
reserved_drafts 589 60 239 30
Total 2,279 689 1,000 332

The release contains 302 training sessions and 128 test sessions, with 4,558 training image references and 2,000 test image references. Every released image reference has a corresponding local image file.

Directory Layout

osworld-persona/
  README.md
  automation_first/
    metadata.json
    train/
      annotations.json
      images/<session_id>/*.png
    test/
      annotations.json
      images/<session_id>/*.png
  eager_handson/
  learning_oriented/
  reserved_drafts/

Each persona directory follows the same layout.

Persona Metadata

Each metadata.json contains:

Field Description
source Source environment and trajectory version.
persona Stable machine-readable persona ID.
ground_truth_persona Full preference profile used to synthesize labels.
annotator Model used to annotate the decision points.
statistics Total and positive examples for each split.

The ground-truth profile is intended for the oracle-persona condition. Methods that induce a persona should infer it only from permitted training history and must not read ground_truth_persona.

Annotation Schema

annotations.json is a JSON list. Every element contains:

Field Type Description
images list of strings Screenshot paths relative to the split directory, in temporal order.
timestamps list of strings Elapsed mm:ss timestamps aligned one-to-one with images.
session string OSWorld session identifier.
summarized_observation string Summary of the visible state and recent activity.
user_intent string or list of strings Inferred immediate user intent; a list represents multiple concurrent intents.
want_support "yes" or "no" Whether this persona would welcome proactive support now.
rationale string Explanation for the preference-conditioned label.
required_context_for_support list of strings Context an assistant would need to provide useful help.

Example (shortened for readability):

{
  "images": [
    "images/076/step_17_20260529@155545467102.png",
    "images/076/step_19_20260529@155608872495.png"
  ],
  "timestamps": ["01:31", "01:54"],
  "session": "076",
  "summarized_observation": "The user is reviewing extracted presentation notes in a terminal.",
  "user_intent": "Understand all requested slide changes before editing the presentation.",
  "want_support": "yes",
  "rationale": "The task contains several repetitive mechanical edits that this persona prefers to automate.",
  "required_context_for_support": [
    "The extracted speaker notes and requested changes"
  ]
}

Licensing Information

This dataset release is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You may share and adapt the material for any purpose, provided that appropriate credit is given.

Citation

Please cite our paper if you use this dataset or the data synthesis pipeline in your work:

@article{shao2026deep,
  title={Deep Personalization: Continuous Context Observation for Local Proactive Agents},
  author={Shao, Yijia and Bian, Song and Guo, Junxian and Jiang, Jindong and Dong, Zhen and Mao, Huizi and Yang, Diyi and Han, Song},
  journal={arXiv preprint arXiv:2611.13457},
  year={2026}
}
Downloads last month
22

Paper for collaborative-agents/OSWorld-Persona