DigiData

Overview DigiData is a dataset designed to offer diverse and high-quality data to train mobile control agents. Different from existing datasets, DigiData is created using a data collection protocol that attempts to comprehensively cover all app features, while simultaneously ensuring high data quality. Repository:** https://github.com/facebookresearch/digidata Paper:** https://arxiv.org/abs/2511.07413 Data Fields **`episode_id`** (**str**): Unique identifier for each trajectory **`step_id`** (**int**): Zero-indexed current position in the episode **`episode_len`** (**int**): Total length of the episode (does not reflect missing steps) **`app`** (**str**): The Android app necessary to complete the goal **`action`** (**str**): The action and its parameters. Possible actions: - `type( )` - `tap(x,y)` - `swipe(x1,y1,x2,y2)` - `navigate({back, home, enter})` - `status({complete, impossible})` **`goal`** (**str**): Text description of the task the agent is expected to perform in the current episode **`action_history`** (**list[str]**): List of previous actions taken at prior steps **`xml`** (**str**): Path to the XML file **`image`** (**str**): Path to the image file of the screen at the current step **`image_history`** (**list[str]**): Paths to images at previous steps **`complete`** (**bool**): Whether a step is missing from the episode in the JSONL file **`eval_category`** (**str**): Evaluation category for the goal: `SEEN`, `NOVEL`, or `FAMILIAR` **`conversations`** (**list[dict]**): The prompt provided to the model and the model's expected response (the action for the next step) Data Splits We have train and test splits in `digidata_train.jsonl` and `digidata_test.jsonl` respectively. We also have COT-annotated splits located in `cot_digidata_train.jsonl` and `cot_digidata_test.jsonl`. Instructions to Use Script to download and extract the DigiData dataset from Hugging Face. Downloads tar.gz archives of images and XML files from the `facebook/DigiData` dataset repo,