Update README.md
#4
by
Tonic
- opened
README.md
CHANGED
@@ -40,3 +40,205 @@ configs:
|
|
40 |
- split: train
|
41 |
path: data/train-*
|
42 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
- split: train
|
41 |
path: data/train-*
|
42 |
---
|
43 |
+
|
44 |
+
# Dataset Card: Android Control Episodes (with Screenshots)
|
45 |
+
|
46 |
+
## Dataset Summary
|
47 |
+
|
48 |
+
This dataset contains Android UI control episodes consisting of:
|
49 |
+
|
50 |
+
- episode-level metadata (`episode_id`, `goal`)
|
51 |
+
- step-by-step instructions (`step_instructions`)
|
52 |
+
- action sequences (`actions` as a list of structs)
|
53 |
+
- base64-encoded screenshots per step (`screenshots_b64`)
|
54 |
+
|
55 |
+
Each episode records a short interaction trajectory on an Android device, including what the agent/user attempted to do and how (tap, swipe, text input, etc.).
|
56 |
+
|
57 |
+
## Supported Tasks and Benchmarks
|
58 |
+
|
59 |
+
- UI task planning and control
|
60 |
+
- Multimodal grounding (text instructions + UI visual context)
|
61 |
+
- Imitation learning / behavior cloning for mobile agents
|
62 |
+
- Action prediction and trajectory modeling
|
63 |
+
|
64 |
+
## Languages
|
65 |
+
|
66 |
+
- Prompts and instructions: English
|
67 |
+
|
68 |
+
## Dataset Structure
|
69 |
+
|
70 |
+
### Data Fields
|
71 |
+
|
72 |
+
- `episode_id` (int64): Unique identifier of the episode.
|
73 |
+
- `goal` (string): Natural language description of the objective for the episode.
|
74 |
+
- `screenshots_b64` (list[string]): Base64-encoded screenshots captured along the trajectory. Large; dominates file sizes.
|
75 |
+
- `actions` (list[struct]): Sequence of actions taken. Each element has:
|
76 |
+
- `action_type` (string): e.g., "open_app", "click", "swipe", "type".
|
77 |
+
- `app_name` (string or null): App associated with the action, if any.
|
78 |
+
- `direction` (string or null): For gestures like swipe (e.g., "up", "down").
|
79 |
+
- `text` (string or null): Text content for typing actions, if applicable.
|
80 |
+
- `x` (int64 or null): X coordinate for tap/click.
|
81 |
+
- `y` (int64 or null): Y coordinate for tap/click.
|
82 |
+
- `step_instructions` (list[string]): Short imperative instructions per step.
|
83 |
+
|
84 |
+
### Example Instance (images excluded for brevity)
|
85 |
+
|
86 |
+
```json
|
87 |
+
{
|
88 |
+
"episode_id": 13,
|
89 |
+
"goal": "On cruisedeals, I would like to view the cruise schedules for a four-night trip from New York to Canada.",
|
90 |
+
"actions": [
|
91 |
+
{"action_type": "open_app", "app_name": "CruiseDeals", "direction": null, "text": null, "x": null, "y": null},
|
92 |
+
{"action_type": "click", "app_name": null, "direction": null, "text": null, "x": 313, "y": 742},
|
93 |
+
{"action_type": "swipe", "app_name": null, "direction": "up", "text": null, "x": null, "y": null}
|
94 |
+
],
|
95 |
+
"step_instructions": [
|
96 |
+
"Open the cruisedeals app",
|
97 |
+
"Click on the suggested searched result",
|
98 |
+
"Swipe up to view schedules"
|
99 |
+
],
|
100 |
+
"screenshots_b64": ["<base64>", "<base64>", "<base64>"]
|
101 |
+
}
|
102 |
+
```
|
103 |
+
|
104 |
+
## Data Splits
|
105 |
+
|
106 |
+
- `train`: 12,232 episodes across 275 Parquet shards (1 row group per file)
|
107 |
+
- `test`: 3,051 episodes across 67 Parquet shards (1 row group per file)
|
108 |
+
|
109 |
+
Total compressed size on the Hub is approximately 67.4 GB (train ≈ 54.3 GB, test ≈ 13.1 GB). The `screenshots_b64` column contributes the majority of the size.
|
110 |
+
|
111 |
+
Typical per-shard stats (example shard):
|
112 |
+
|
113 |
+
- ~45 episodes per shard
|
114 |
+
- ~6–7 screenshots per episode on average
|
115 |
+
- ~5–6 actions per episode on average
|
116 |
+
- ~5–6 step instructions per episode on average
|
117 |
+
|
118 |
+
## Usage
|
119 |
+
|
120 |
+
### Load with Datasets (streaming to avoid full download)
|
121 |
+
|
122 |
+
```python
|
123 |
+
from datasets import load_dataset
|
124 |
+
|
125 |
+
ds = load_dataset(
|
126 |
+
"parquet",
|
127 |
+
data_files="hf://datasets/<owner>/<repo>@~parquet/default/train/*.parquet",
|
128 |
+
streaming=True,
|
129 |
+
)["train"]
|
130 |
+
|
131 |
+
for i, ex in enumerate(ds):
|
132 |
+
ex.pop("screenshots_b64", None) # skip large images for lightweight inspection
|
133 |
+
print(ex["episode_id"], ex["goal"])
|
134 |
+
if i >= 4:
|
135 |
+
break
|
136 |
+
```
|
137 |
+
|
138 |
+
### Materialize a small slice without streaming
|
139 |
+
|
140 |
+
```python
|
141 |
+
from datasets import load_dataset
|
142 |
+
|
143 |
+
small = load_dataset(
|
144 |
+
"parquet",
|
145 |
+
data_files="hf://datasets/<owner>/<repo>@~parquet/default/train/*.parquet",
|
146 |
+
split="train[:1%]",
|
147 |
+
)
|
148 |
+
print(len(small))
|
149 |
+
```
|
150 |
+
|
151 |
+
### DuckDB: schema preview and lightweight sampling
|
152 |
+
|
153 |
+
```python
|
154 |
+
import duckdb
|
155 |
+
|
156 |
+
# Peek schema of one shard
|
157 |
+
duckdb.sql("""
|
158 |
+
DESCRIBE SELECT * FROM
|
159 |
+
'hf://datasets/<owner>/<repo>@~parquet/default/train/0000.parquet'
|
160 |
+
""").show()
|
161 |
+
|
162 |
+
# Count rows via metadata only (no full scan)
|
163 |
+
duckdb.sql("""
|
164 |
+
SELECT SUM(row_group_num_rows) AS total_rows
|
165 |
+
FROM parquet_metadata('hf://datasets/<owner>/<repo>@~parquet/default/train/*.parquet')
|
166 |
+
""").show()
|
167 |
+
|
168 |
+
# Sample a few rows excluding heavy images
|
169 |
+
duckdb.sql("""
|
170 |
+
SELECT episode_id, goal,
|
171 |
+
list_length(actions) AS num_actions,
|
172 |
+
list_length(step_instructions) AS num_steps
|
173 |
+
FROM 'hf://datasets/<owner>/<repo>@~parquet/default/train/*.parquet'
|
174 |
+
LIMIT 10
|
175 |
+
""").show()
|
176 |
+
```
|
177 |
+
|
178 |
+
### PyArrow: footer-only metadata or row-group reads
|
179 |
+
|
180 |
+
```python
|
181 |
+
from huggingface_hub import HfFileSystem
|
182 |
+
import pyarrow.parquet as pq
|
183 |
+
|
184 |
+
fs = HfFileSystem()
|
185 |
+
path = "hf://datasets/<owner>/<repo>@~parquet/default/train/0000.parquet"
|
186 |
+
|
187 |
+
# Metadata-only: schema & row groups
|
188 |
+
with fs.open(path, "rb") as f:
|
189 |
+
pf = pq.ParquetFile(f)
|
190 |
+
print(pf.schema_arrow)
|
191 |
+
print(pf.metadata.num_rows, pf.num_row_groups)
|
192 |
+
|
193 |
+
# Read a single row group without images
|
194 |
+
with fs.open(path, "rb") as f:
|
195 |
+
pf = pq.ParquetFile(f)
|
196 |
+
cols = [c for c in pf.schema_arrow.names if c != "screenshots_b64"]
|
197 |
+
tbl = pf.read_row_group(0, columns=cols)
|
198 |
+
print(tbl.slice(0, 3).to_pydict())
|
199 |
+
```
|
200 |
+
|
201 |
+
### Dask: predicate/projection pushdown
|
202 |
+
|
203 |
+
```python
|
204 |
+
import dask.dataframe as dd
|
205 |
+
|
206 |
+
ddf = dd.read_parquet(
|
207 |
+
"hf://datasets/<owner>/<repo>@~parquet/default/train/*.parquet",
|
208 |
+
columns=["episode_id", "goal", "actions", "step_instructions"],
|
209 |
+
)
|
210 |
+
print(ddf.head())
|
211 |
+
```
|
212 |
+
|
213 |
+
## Efficiency Tips
|
214 |
+
|
215 |
+
- Prefer streaming or column selection to avoid downloading `screenshots_b64` unless needed.
|
216 |
+
- Use DuckDB `parquet_metadata(...)` or PyArrow `ParquetFile(...).metadata` to inspect sizes/counts without reading data pages.
|
217 |
+
- Each file has one row group; shard-level parallelism is straightforward.
|
218 |
+
|
219 |
+
## Licensing
|
220 |
+
|
221 |
+
[More Information Needed]
|
222 |
+
|
223 |
+
## Citation
|
224 |
+
|
225 |
+
If you use this dataset in your work, please cite the source dataset/creators as appropriate and this repository. Example placeholder:
|
226 |
+
|
227 |
+
```bibtex
|
228 |
+
@misc{android_control_episodes,
|
229 |
+
title = {Android Control Episodes Dataset},
|
230 |
+
year = {2025},
|
231 |
+
url = {https://huggingface.co/datasets/smolagents/android-control}
|
232 |
+
}
|
233 |
+
```
|
234 |
+
|
235 |
+
## Limitations and Risks
|
236 |
+
|
237 |
+
- Screenshots are stored as base64 strings and can be large; consider storage and memory implications.
|
238 |
+
- Some action fields (e.g., `app_name`, `direction`, `text`) may be null for many steps.
|
239 |
+
- Visual UI elements may vary across Android versions/devices.
|
240 |
+
|
241 |
+
## Maintainers
|
242 |
+
|
243 |
+
[more information needed]
|
244 |
+
|