OpenMobile-Data-v2 / README.md
yanhhh's picture
Upload README.md with huggingface_hub
992f42e verified
|
Raw History Blame Contribute Delete
3.27 kB
---
pretty_name: OpenMobile-Data-v2
language:
- en
- zh
task_categories:
- image-text-to-text
tags:
- mobile
- gui-agent
- multimodal
- emulator
- simulator
size_categories:
- 100K<n<1M
configs:
- config_name: sft
data_files:
- split: train
path:
- emulator/train.json
- simulator/gui-only/train.json
- simulator/hybrid/train.json
- config_name: rl
data_files:
- split: train
path: RL/task.jsonl
---
# OpenMobile-Data-v2
Two parts. SFT is one decision step per example: the model reads the remaining screenshots and the conversation history, then writes a `Thought:` followed by one tool call. Coordinates belong in the tool arguments. The screen space in the prompt is **999×999**. RL is one task per line, with the instruction, the initial state, and the check for success.
SFT has **11,643** trajectories and **118,582** steps, about **59.2 GB**. Trajectory folders contain only `screenshot_step{K}.png`. RL has **2,475** tasks in one JSONL file.
## Splits
| Split | File | Trajectories | Steps | Tools |
| --- | --- | ---: | ---: | --- |
| Emulator | `emulator/train.json` | 7,725 | 83,631 | `mobile_use` only. 3,202 cross-app, 4,523 single-app |
| Simulator, GUI-only | `simulator/gui-only/train.json` | 2,144 | 22,857 | `mobile_use` only |
| Simulator, hybrid | `simulator/hybrid/train.json` | 1,774 | 12,094 | `mobile_use`, plus extra app tools listed in the current tool message |
`dataset_info.json` registers these files as `emulator_qwen35_last3_noloop_notitle_nocoord_think_thought`, `mobilegym_qwen35_last3_noloop_notitle_nocoord_think_thought`, and `gui_mcp_qwen35_last3_noloop_notitle_nocoord_think_thought`.
## Record format
Each `train.json` is a list of objects with `id`, `messages`, and `images`.
- `images` holds at most the last three screenshots, oldest first. Paths are relative to that JSON file.
- Earlier screenshots are replaced in the text by `This screenshot has been collapsed.` The remaining `<image>` placeholders line up with `images`.
- History `assistant` turns are marked `"loss": false`. Supervise the last `assistant` turn.
- Hybrid `tool` messages may include a foreground-app tool catalog. Those tools are valid only while that app stays in front.
## RL tasks
`RL/tasks.jsonl` is one JSON object per line. Each line has `task`, `init`, and `verify` together.
- `task.instruction` is the user goal. `task.apps` lists the apps involved.
- `init` is the starting state, including `assertion_initials`.
- `verify` is the success check: `assertions` and optional `extra_checks`.
## Use with ms-swift
Download the full repository so the screenshot paths resolve. The dataset viewer table stores those paths as strings and does not replace this download.
```bash
hf download OpenMobile-2/OpenMobile-Data-v2 --repo-type dataset --local-dir OpenMobile-Data-v2
swift sft \
--custom_dataset_info OpenMobile-Data-v2/dataset_info.json \
--dataset emulator_qwen35_last3_noloop_notitle_nocoord_think_thought \
mobilegym_qwen35_last3_noloop_notitle_nocoord_think_thought \
gui_mcp_qwen35_last3_noloop_notitle_nocoord_think_thought
```
Adjust `model`, batch size, and output paths for your local [ms-swift](https://github.com/modelscope/ms-swift) setup and hardware.