Blame for README.md

---
pretty_name: LGWM data and RSWT-Bench
language:
- en
size_categories:
- 1M<n<10M
tags:
- world-model
- gui-agent
- computer-use
- mobile-agent
- agent-safety
- android
configs:
- config_name: transitions
  default: true
  data_files:
  - split: train
    path: index/*/train/*.parquet
  - split: validation
    path: index/*/validation/*.parquet
- config_name: rswt_bench
  data_files:
  - split: test
    path: benchmark/rswt_bench_test.parquet
  - split: train
    path: benchmark/rswt_bench_train.parquet
---

# LGWM data and RSWT-Bench

Data for **Right Screen, Wrong Transition: World Models as Verifiers for GUI Agents**:
the processed GUI transitions used to train [LGWM](https://huggingface.co/jiamingzz/lgwm)
and the RSWT-Bench benchmark.

[Project page](https://jiamingzhang.netlify.app/lgwm/) · [Code](https://github.com/jiamingzhang94/lgwm) · [Model weights](https://huggingface.co/jiamingzz/lgwm) · [Paper](https://arxiv.org/abs/2610.11942)

## Transitions

Each transition is a before screenshot, an action and an after screenshot.

| Source | Train | Validation |
| --- | ---: | ---: |
| [Android in the Wild](https://github.com/google-research/google-research/tree/master/android_in_the_wild) | 639,758 | 13,392 |
| [MiniWoB++](https://github.com/Farama-Foundation/miniwob-plusplus) | 986,151 | – |
| [GUIOdyssey](https://huggingface.co/datasets/OpenGVLab/GUI-Odyssey) | 95,788 | 5,045 |
| [AndroidControl](https://github.com/google-research/google-research/tree/master/android_control) | 70,514 | 670 |
| [AMEX](https://huggingface.co/datasets/Yuxiang007/AMEX) | 32,613 | 1,728 |
| Total | 1,824,824 | 20,835 |

| Path | Content |
| --- | --- |
| `processed/` | Screenshot shards (256×512 WebP) with actions, 36 GB |
| `index/` | One row per transition: shard, byte offsets of the two screenshots, action |
| `text_table.npz` | Action-text embeddings |

The viewer shows the `index/` tables. In them, `action_type` is 0 tap, 1 long_press,
2 swipe, 3 scroll, 4 type_text, 5 key, 6 open_app, 7 wait, 8 click, 9 hover, 10 drag;
`x`, `y`, `x2`, `y2` are normalized to 0–1; `direction` is 0 up, 1 down, 2 left, 3 right.

## RSWT-Bench

RSWT-Bench (Right Screen, Wrong Transition) is a donor-paired benchmark for
transition verification. In each of 222 donor pairs, a credential screen appears once as a
legitimate outcome and once, pixel-identical, as a hijacked one, so any detector
that sees only the resulting screen scores 0.5 AUC by construction. The test split
has 966 transitions; the train split has 6,786 for supervised methods.

| Field | Meaning |
| --- | --- |
| `base` | Transition supplying the current screen and action |
| `obs` | Transition supplying the observed next screen |
| `pair_id` | Hijacked/paired-safe example pair |
| `donor_id` | Replacement-screen identity used for the paired bootstrap |
| `cls` | 1 hijacked, 2 paired safe, 3 legitimate credential, 4 surprising benign, 5 benign mismatch |

## Usage

```bash
git clone https://github.com/jiamingzhang94/lgwm && cd lgwm
pip install -c requirements-tested.txt -e '.[text,training,evaluation]'
hf download jiamingzz/lgwm --local-dir weights
hf download jiamingzz/lgwm-data --repo-type dataset --local-dir data

python tools/extract_features.py --manifest benchmark/rswt_bench_test.jsonl \
  --data-root data --index-dir data/index --output outputs/features/test
python tools/evaluate_benchmark.py --features outputs/features/test \
  --output outputs/eval/test.json
```

## Citation

```bibtex
@article{zhang2026lgwm,
  title   = {Right Screen, Wrong Transition:
             World Models as Verifiers for GUI Agents},
  author  = {Zhang, Jiaming and Wang, Xuan and Zhang, Fuyao and
             Cao, Yang and Lyu, Lingjuan and Lim, Wei Yang Bryan},
  journal = {arXiv preprint arXiv:2610.11942},
  year    = {2026}
}
```