GroundGUI_RL: 12,600 RL grounding prompts derived from GroundCUA (MIT)

LICENSE

@@ -0,0 +1,31 @@
+GroundGUI_RL is derived entirely from GroundCUA
+(https://huggingface.co/datasets/ServiceNow/GroundCUA), which is released under
+the MIT License. The screenshots, instructions and bounding boxes come from
+GroundCUA; this release only selects, re-mixes and re-formats them. It is
+distributed under the same MIT License.
+
+MIT License
+
+Copyright (c) The GroundCUA Team
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+Screenshots depict third-party desktop applications whose interfaces remain
+the property of their respective owners. Users are responsible for ensuring
+compliance with all applicable laws and policies.

README.md

@@ -0,0 +1,134 @@
+---
+license: mit
+language:
+  - en
+task_categories:
+  - image-to-text
+tags:
+  - computer_use
+  - gui-grounding
+  - reinforcement-learning
+  - GroundCUA
+size_categories:
+  - 10K<n<100K
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: data/train-*.parquet
+---
+
+# GroundGUI_RL
+
+A 12,600-prompt reinforcement-learning corpus for desktop GUI grounding. Each
+prompt is a screenshot plus a natural-language instruction, and the reward is
+whether the predicted click falls inside the target element's bounding box.
+
+**All content comes from [GroundCUA](https://huggingface.co/datasets/ServiceNow/GroundCUA)**
+(MIT). The screenshots are byte-identical to GroundCUA's, and the instructions
+and boxes are GroundCUA's annotations. This release adds only sample
+selection, slice mixing and formatting into a Qwen3-VL `computer_use`
+tool-call prompt. Please cite GroundCUA (below) if you use it.
+
+| | |
+|---|---|
+| Prompts | 12,600 |
+| Screenshots | 11,205 (at most 2 prompts per image) |
+| Applications | 87 desktop apps (from GroundCUA) |
+| Size | ~4.9 GB (10 parquet shards, images embedded) |
+
+## Layout
+
+```
+GroundGUI_RL/
+├── data/train-0000{0..9}-of-00010.parquet   # 12,600 rows, screenshots embedded as image bytes
+├── SHA256SUMS
+├── LICENSE
+└── README.md
+```
+
+## Record format
+
+```text
+{
+  "images": [<PIL.Image 1920x1008>],
+  "messages": [
+    {"role": "system",    "content": "<computer_use tool definition; screen given as 1000x1000>"},
+    {"role": "user",      "content": "<image>\nSelect a gradient to apply a smooth color transition ..."},
+    {"role": "assistant", "content": "<tool_call>\n{\"name\": \"computer_use\", \"arguments\": {\"action\": \"left_click\", \"coordinate\": [96, 399]}}\n</tool_call>"}
+  ],
+  "solution": {"name": "mobile_use", "arguments": {"action": "click", "coordinate": [139, 381, 232, 426]}},
+  "additional_paras": "{\"image_size\": [1920, 1008], \"app\": \"Krita\", \"ui_type\": \"icon\", ...}",
+  "sample_id": 4026033
+}
+```
+
+Coordinate conventions (note that the two fields use different units):
+
+- `solution.arguments.coordinate` is the target box `[x1, y1, x2, y2]` in
+  **absolute pixels** of the original image. Reward functions should
+  normalise it with `image_size`.
+- The assistant turn is a reference answer: the box centre in **0–1000
+  normalised** coordinates (Qwen3-VL convention). It is useful for SFT
+  warm-up. An RL reward should use `solution`.
+- `additional_paras` is a **JSON string**, not an object.
+
+`additional_paras` fields:
+
+| field | meaning |
+|---|---|
+| `image_size` | `[width, height]` in pixels (matches the PNG header for every row) |
+| `app` | GroundCUA application name |
+| `ui_type` | `icon` or `text` element |
+| `register` | `functional_instruction`, `spatial_relation` or `bare_label` |
+| `source` | which GroundCUA annotation file the instruction came from |
+| `elem_text` | GroundCUA element text of the target |
+| `area_frac` | target box area / image area |
+| `slice` | mixing slice (below) |
+
+## Slices
+
+| slice | rows | content |
+|---|---:|---|
+| `base_icon_func` | 4,848 | GroundCUA functional instructions targeting icon elements |
+| `base_text_func` | 1,500 | GroundCUA functional instructions targeting text elements |
+| `icon_func_new` | 1,652 | additional functional instructions on icon elements |
+| `bare_label` | 2,000 | the element's own label as the instruction, e.g. "Wrap Text." (mostly text-labelled buttons) |
+| `spatial` | 2,000 | GroundCUA spatial-relation instructions, e.g. "... immediately right of \"app icon\"" (about 79% horizontal) |
+| `small` | 600 | small targets, 80–250 ppm of the image area |
+
+## Usage
+
+```python
+import json
+from datasets import load_dataset
+
+ds = load_dataset("tjcaspar/GroundGUI_RL", split="train")
+row = ds[0]
+img = row["images"][0]                         # PIL.Image
+extra = json.loads(row["additional_paras"])    # image_size, slice, ...
+x1, y1, x2, y2 = row["solution"]["arguments"]["coordinate"]   # pixels
+```
+
+ms-swift (GRPO / RLOO) can take the hub id directly:
+
+```bash
+swift rlhf --rlhf_type grpo --dataset tjcaspar/GroundGUI_RL --reward_funcs <your-bbox-hit-reward> ...
+```
+
+## License and attribution
+
+MIT, inherited from GroundCUA (see `LICENSE`). Screenshots show third-party
+applications whose interfaces remain the property of their owners.
+
+```bibtex
+@misc{feizi2025groundingcomputeruseagents,
+      title={Grounding Computer Use Agents on Human Demonstrations},
+      author={Aarash Feizi and Shravan Nayak and Xiangru Jian and Kevin Qinghong Lin and Kaixin Li and Rabiul Awal and Xing Han Lù and Johan Obando-Ceron and Juan A. Rodriguez and Nicolas Chapados and David Vazquez and Adriana Romero-Soriano and Reihaneh Rabbany and Perouz Taslakian and Christopher Pal and Spandana Gella and Sai Rajeswar},
+      year={2025},
+      eprint={2511.07332},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG},
+      url={https://arxiv.org/abs/2511.07332},
+}
+```

SHA256SUMS

@@ -0,0 +1,10 @@
+e5b1832d6d59afb1f92540b70d71f13a156bddfd94d9930c1fa1c22c8bbc4a5b  data/train-00000-of-00010.parquet
+a0154cc7f85f4685a87447d64bfc24fb3925e3b0bbdbbd55e7fcf72c3f72875a  data/train-00001-of-00010.parquet
+9a1966921a5cd120f7a9a50570463cd00d896e5f63eaf2b8a360f814d841f199  data/train-00002-of-00010.parquet
+73bef7dd8e4e5d40f71c96807f94a5595d39d57c97427f4e579470dc8b28a175  data/train-00003-of-00010.parquet
+c1f3e2fa816027bda710fc489865d8ff09395c63a352cc6e7376c8858bae0d23  data/train-00004-of-00010.parquet
+7ae4c592b3ddc2c9e49c28e4360738053967af29140bb4fca3e37c8fb2be8936  data/train-00005-of-00010.parquet
+b3b3f71f172df9e423503cb26cbf0f26f19e7ca06301ef4145c0012f68a7e3fa  data/train-00006-of-00010.parquet
+1bb756ed2fbe6d074c46b63fa8666a1c67899096a6af2ab12a05bef817727b6f  data/train-00007-of-00010.parquet
+583b3f2cdf983b158650ea4d0e12e31779fa7a05506939c3f750bc1be0a902d0  data/train-00008-of-00010.parquet
+04b9ec230912fdf9b5f51e0bd1d8d3b266e71c077b3853dd7fd4d729230b9ad3  data/train-00009-of-00010.parquet

data/train-00000-of-00010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5b1832d6d59afb1f92540b70d71f13a156bddfd94d9930c1fa1c22c8bbc4a5b
+size 502289364

data/train-00001-of-00010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0154cc7f85f4685a87447d64bfc24fb3925e3b0bbdbbd55e7fcf72c3f72875a
+size 524121212

data/train-00002-of-00010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a1966921a5cd120f7a9a50570463cd00d896e5f63eaf2b8a360f814d841f199
+size 505991322

data/train-00003-of-00010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73bef7dd8e4e5d40f71c96807f94a5595d39d57c97427f4e579470dc8b28a175
+size 531076977

data/train-00004-of-00010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1f3e2fa816027bda710fc489865d8ff09395c63a352cc6e7376c8858bae0d23
+size 518169442

data/train-00005-of-00010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ae4c592b3ddc2c9e49c28e4360738053967af29140bb4fca3e37c8fb2be8936
+size 531973653

data/train-00006-of-00010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3b3f71f172df9e423503cb26cbf0f26f19e7ca06301ef4145c0012f68a7e3fa
+size 519333962

data/train-00007-of-00010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bb756ed2fbe6d074c46b63fa8666a1c67899096a6af2ab12a05bef817727b6f
+size 527248806

data/train-00008-of-00010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:583b3f2cdf983b158650ea4d0e12e31779fa7a05506939c3f750bc1be0a902d0
+size 483386531

data/train-00009-of-00010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04b9ec230912fdf9b5f51e0bd1d8d3b266e71c077b3853dd7fd4d729230b9ad3
+size 532226217