CoolFace
Datasetpublic

sharryXR/MobileWorld-ProgRM-Rollouts

MobileWorld ProgRM Rollouts Private training release containing multimodal ProgRM and ORM data derived from MobileWorld rollouts. The release is intended for reward-model training and subsequent reinforcement-learning research. Release Version: v1.0.0 ProgRM code commit: 875ba94c74a9742e903acaa81895f5685de7eac1 on branch mobileworld Tasks: 161 Canonical trajectories: 966 Splits: {"test": 161, "train": 644, "val": 161} Screenshot shards: 11 Run… See the full description on the dataset page: https://huggingface.co/datasets/sharryXR/MobileWorld-ProgRM-Rollouts.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face
0likes253downloads
Dataset Card

MobileWorld ProgRM Rollouts

Private training release containing multimodal ProgRM and ORM data derived from MobileWorld rollouts. The release is intended for reward-model training and subsequent reinforcement-learning research.

Release

  • —Version: v1.0.0
  • —ProgRM code commit: 875ba94c74a9742e903acaa81895f5685de7eac1 on branch mobileworld
  • —Tasks: 161
  • —Canonical trajectories: 966
  • —Splits: {"test": 161, "train": 644, "val": 161}
  • —Screenshot shards: 11

Run scripts/prepare_mobileworld_hf_dataset.py from the ProgRM repository to download, verify, extract, and materialize absolute image paths for MS-SWIFT.

Contents

  • —data/progrm: step-level progress-regression JSONL.
  • —data/orm: trajectory-level outcome-regression JSONL.
  • —manifests: selection, rejection, statistics, and data-integrity metadata.
  • —source_trajectories.jsonl: canonical trajectories with runtime logs and token metadata removed.
  • —shards: deterministic tar.zst archives containing raw screenshots.

Each training row follows the MS-SWIFT multimodal sequence-classification format with messages, one relative path in images, and a scalar label.

Labels and limitations

Successful ProgRM trajectories use linear labels from 0 to 1. Failed trajectories are labelled 0 at every step because the rollout only provides a terminal binary evaluator result. ORM labels are terminal success values. Forty-eight tasks have only failed trajectories and three have only successful trajectories. Ask-user values and UI contents are generated inside the MobileWorld benchmark; they may resemble email addresses or phone numbers. The dataset is private and must be reviewed before any public release.

License

MobileWorld is distributed under Apache-2.0. This derived rollout release uses the same license and should retain attribution to the MobileWorld project.