Datasets:
File size: 4,657 Bytes
c7d7f6a e08144f c7d7f6a 1ba6f57 c7d7f6a bf40bdc c7d7f6a bf40bdc 3d8d057 bf40bdc 3d8d057 c7d7f6a cbf853b 1ba6f57 cbf853b 1ba6f57 cbf853b c7d7f6a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 | ---
license: cc-by-4.0
task_categories:
- image-text-to-text
- visual-question-answering
language:
- en
- zh
pretty_name: InquireMobile
size_categories:
- 10K<n<100K
tags:
- mobile-agent
- gui-agent
- vlm
- rl
- interactive
---
# InquireMobile
Dataset for **InquireMobile: Teaching VLM-based Mobile Agent to Request Human
Assistance via Reinforcement Fine-Tuning** (ACL 2026).
InquireMobile is an interactive mobile GUI agent that *actively seeks human
confirmation at critical decision points* instead of acting autonomously. This
dataset contains:
- **InquireBench** evaluation tasks (instruction + category + intent), and
- the **training trajectories** used for supervised fine-tuning (SFT) and
reinforcement fine-tuning (GRPO / RLVR).
> Project page: <https://bit-aqh.github.io/InquireMobile/homepage/>
> Paper: <https://aclanthology.org/2026.acl-long.1487/>
## Contents
The dataset is organized as:
- `data/` — all JSONL data files (see below).
- `images/` — screenshots referenced by the data files (see [Images](#images)).
All data files are line-delimited JSON (JSONL) unless noted otherwise.
### Evaluation tasks (InquireBench)
| File | Description |
|---|---|
| `tasks_en.jsonl`, `tasks_zh.jsonl` | Core evaluation task instructions (English / Chinese). |
| `tasks_en_extended.jsonl` | Extended English task set. |
| `tasks_{en,zh}_huawei.jsonl`, `tasks_{en,zh}_simulator.jsonl` | Task subsets for the Huawei device / simulator environments. |
| `tasks_zh_huawei_7.jsonl` | Huawei-device subset (7 tasks). |
Task schema:
```json
{
"instruction": "Open WeChat, browse my Moments ...",
"apps": ["wechat"],
"category": "privacy",
"language": "en",
"need_login": true,
"intension": "Open WeChat, browse my Moments ..."
}
```
### Training data
| File | Description |
|---|---|
| `interactive_sft.jsonl` | Stage-1 SFT trajectories (conversation format). |
| `interactive_grpo_401408.jsonl` | Stage-2 GRPO / RLVR trajectories (401408 max-pixels). |
| `general_grpo_5k.jsonl` | General GRPO training data (5k). |
| `sft_general.jsonl` | SFT data (general). |
| `sft_pos.jsonl` | SFT data (positive samples). |
### Instruction pool
| File | Description |
|---|---|
| `instructions.jsonl` | Raw instruction pool (Chinese + English), with optional `image` links to the instruction screenshots. |
GRPO / trajectory schema:
```json
{
"goal": "...",
"action": "...",
"bbox": [x1, y1, x2, y2],
"width": 1080,
"height": 1920,
"resized_width": 432,
"resized_height": 768,
"img_list": ["..."],
"system": "...",
"query": "..."
}
```
SFT schema:
```json
{
"conversations": [
{"from": "system", "value": "..."},
{"from": "human", "value": "<image>..."},
{"from": "gpt", "value": "..."}
],
"images": ["..."]
}
```
## Images
Images are stored under `images/` and are referenced by **relative paths** from
the data files (relative to the dataset root).
| Directory | Description |
|---|---|
| `images/trajectory/` | Stage-2 GRPO rollout screenshots (UUID-named), referenced by `img_list` in the GRPO data. |
| `images/instruction/` | The 72 English instruction screenshots (`1.png`…`72.png`), referenced by the `images` field of the SFT data and by `interactive_added.jsonl`. |
The data files reference images by their basename (UUID) or number (`N.png`); map
these to the directories above when reproducing training/evaluation.
## Usage
```python
import json
with open("data/tasks_en.jsonl", encoding="utf-8") as f:
tasks = [json.loads(line) for line in f]
```
Or download directly from Hugging Face:
```bash
huggingface-cli download PG23/InquireMobile --repo-type dataset --local-dir ./InquireMobile-data
```
## License
- **Data**: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)
## Citation
```bibtex
@inproceedings{ai-etal-2026-inquiremobile,
title = "{I}nquire{M}obile: Teaching {VLM}-based Mobile Agent to Request Human Assistance via Reinforcement Fine-Tuning",
author = "Ai, Qihang and
Bu, Pi and
Cao, Yue and
Wang, Yingyao and
Gu, Jihao and
Xing, Jingxuan and
Zhu, Zekun and
Jiang, Wei and
Zheng, Zhicheng and
Song, Jun and
Jiang, Yuning",
booktitle = "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.acl-long.1487/",
doi = "10.18653/v1/2026.acl-long.1487",
pages = "32241--32255"
}
```
|