Polish the dataset card, declare the six splits, and license under MIT

Declare train/val/test_id/test_app/test_theme/test_resolution (WebDataset) as the default config, add a transitions config, keep the preview. MIT license. Remove host, path and personal details from the docs and provenance.

LICENSE

@@ -1,18 +1,21 @@
-DeskForge-1M — licence not yet determined
-=========================================
+MIT License
 
-This dataset has NOT been assigned a licence and is NOT cleared for
-redistribution.
+Copyright (c) 2026 The DeskForge authors
 
-The screenshots contain third-party material whose licensing has not been
-reviewed: application user interfaces (GTK applications such as HomeBank,
-Bluefish, Thunar, Chromium and others), icon themes, fonts, desktop wallpapers,
-and, in browser scenes, rendered third-party web pages captured on the date
-shown in the record.
+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:
 
-Until that review is complete and a licence is chosen, this repository is
-private and its contents must not be redistributed, mirrored or published.
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
 
-The annotation files, indexes and code in this repository are the work of the
-dataset author and will be licensed alongside the payload once the review is
-complete.
+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.

README.md

@@ -1,7 +1,6 @@
 ---
 pretty_name: DeskForge-1M
-license: other
-license_name: license-pending-review
+license: mit
 task_categories:
   - image-to-text
   - object-detection
@@ -9,13 +8,44 @@ language:
   - en
 tags:
   - gui
-  - screen-understanding
+  - gui-grounding
+  - screen-parsing
   - computer-use
+  - desktop
   - accessibility
   - webdataset
 size_categories:
   - 1M<n<10M
 configs:
+  - config_name: default
+    default: true
+    data_files:
+      - split: train
+        path: data/train/*.tar
+      - split: val
+        path: data/val/*.tar
+      - split: test_id
+        path: data/test_id/*.tar
+      - split: test_app
+        path: data/test_app/*.tar
+      - split: test_theme
+        path: data/test_theme/*.tar
+      - split: test_resolution
+        path: data/test_resolution/*.tar
+  - config_name: transitions
+    data_files:
+      - split: train
+        path: index/transitions/train.parquet
+      - split: val
+        path: index/transitions/val.parquet
+      - split: test_id
+        path: index/transitions/test_id.parquet
+      - split: test_app
+        path: index/transitions/test_app.parquet
+      - split: test_theme
+        path: index/transitions/test_theme.parquet
+      - split: test_resolution
+        path: index/transitions/test_resolution.parquet
   - config_name: preview
     data_files:
       - split: preview
@@ -24,153 +54,143 @@ configs:
 
 # DeskForge-1M
 
-**1,207,368 annotated Linux desktop screenshots**, every visible element boxed
-from the accessibility tree, with occlusion resolved against the real window
-stack. 135,731 of the scenes are short click explorations, giving 917,211
-state–action–state transitions on top of the still images.
-
-The annotation is not a caption or a bounding box for one target. It is *every*
-element a person can see, with its class, its visible text, its modal geometry
-(the pixels it actually occupies) and its amodal geometry (where it would be if
-nothing covered it).
-
-## What is in it
-
-| split | scenes | episodes | observations | state-eligible | transitions | transition-eligible |
-| --- | ---: | ---: | ---: | ---: | ---: | ---: |
-| `train` | 266,159 | 111,249 | 999,494 | 881,794 | 760,850 | 716,573 |
-| `val` | 2,938 | 1,284 | 11,279 | 10,081 | 8,667 | 8,147 |
-| `test_id` | 5,805 | 2,524 | 22,550 | 20,097 | 17,383 | 16,371 |
-| `test_app` | 11,512 | 6,553 | 43,206 | 39,884 | 32,848 | 29,934 |
-| `test_theme` | 15,390 | 5,586 | 51,627 | 46,730 | 38,671 | 35,636 |
-| `test_resolution` | 21,927 | 8,535 | 79,212 | 69,213 | 58,792 | 55,926 |
-| **total** | **323,731** | **135,731** | **1,207,368** | **1,067,799** | **917,211** | **862,587** |
-
-1,151 uncompressed WebDataset tars, 1.14 TB. 19 applications, 7 desktop themes,
-7 resolutions from 1366×768 to 3840×2160.
-
-**Two eligibility flags, and they are not the same thing.** `state_train_eligible`
-is the recommended filter for the still-image view: publishable, not a
-near-duplicate scene, not a no-op episode frame. `transition_train_eligible` is
-for the action view and **keeps no-op transitions** — a click that changed
-nothing is supervision, not a defect. A no-op frame is excluded from the states
-view while the transition that produced it is kept.
-
-## The evaluation splits are the contribution, not a convenience
-
-A single random test set answers one question — "does it work on data like the
-training data" — so three axes are held out of training **entirely**, not merely
-sampled out:
-
-| axis | held out | asks |
-| --- | --- | --- |
-| `test_app` | `gnome-system-monitor` | an application category never seen |
-| | `pluma`, `xarchiver` | an unseen *instance* of a category training knows |
-| `test_theme` | `quartz_night_nord` | unseen desktop chrome |
-| `test_resolution` | `retina_2880x1800` | unseen scale and aspect ratio |
-
-Holding out an application removes *every* scene containing it, not just the
-test scenes, computed over the union of applications across every frame of a
-scene. Verified zero leakage. See `docs/splits.md` for what each axis cost and
-why these were affordable.
-
-**Occlusion is deliberately not held out.** The corpus argues that this
-supervision improves occlusion robustness, and a model cannot learn that from
-data with the occlusion removed. Occlusion and window count are reported as
-slices of `test_id` instead. They are heavy: median occluded ratio 0.284, p90
-0.743, and more than half of all observations have ≥25% of their elements
-clipped or dropped.
+**DeskForge-1M** is a corpus of **1.21M annotated desktop screenshots** with
+**159.7M element instances** and **917K recorded click transitions**, generated
+with [DeskForge](https://github.com/Saidgurbuz/deskforge), a controllable
+desktop environment that composes and explores real applications.
+
+[**Project page**](https://saidgurbuz.github.io/deskforge/) ·
+[**Code**](https://github.com/Saidgurbuz/deskforge) · **Paper** (coming soon)
+
+A. Said Gurbuz · Ahmed Nassar · Sunghwan Hong · Marc Pollefeys · Peter W. J. Staar
+<br>ETH Zurich · IBM Research Zurich · Microsoft
+
+![DeskForge overview](https://saidgurbuz.github.io/deskforge/assets/img/overview-2400.webp)
+
+Every screenshot is annotated densely: not one target, but **every visible
+element**, with its type, text, geometry, interaction properties, hierarchy and
+owning window. Overlapping windows are resolved, so each element records both
+its full extent and the fragments that are actually visible. Scenes combine
+several real applications with varied content, window layouts, appearance
+presets and display resolutions, and short click explorations link each action
+to the screen before and after it.
+
+## At a glance
+
+| | |
+| --- | --- |
+| Observations | 1,207,368 screenshots in 323,731 scenes |
+| Element instances | 159.7M (132 per screen on average) |
+| Click transitions | 917,211, from 135,731 exploration episodes |
+| Applications | 19 real Linux desktop applications, several per screen |
+| Appearance | 7 presets, from classic Linux to Windows- and macOS-inspired styles |
+| Resolutions | 7, from 1366×768 to 3840×2160 |
+
+## Splits
+
+Splits are made per **scene**, so all frames of a scene stay together. Three test
+splits hold an attribute out of training **entirely**, which measures
+generalization to desktop configurations never seen in training.
+
+| split | held out | observations | transitions |
+| --- | --- | ---: | ---: |
+| `train` | | 999,494 | 760,850 |
+| `val` | | 11,279 | 8,667 |
+| `test_id` | new scenes with seen attributes | 22,550 | 17,383 |
+| `test_app` | GNOME System Monitor, Pluma, Xarchiver | 43,206 | 32,848 |
+| `test_theme` | the Quartz Night Nord preset | 51,627 | 38,671 |
+| `test_resolution` | 2880×1800 | 79,212 | 58,792 |
+
+A held-out application is excluded from every scene that contains it, in any
+frame. See [`docs/splits.md`](docs/splits.md) for how each axis was chosen.
+
+## What a sample contains
+
+The default configuration streams the observations as WebDataset shards. Each
+observation is four members sharing one key:
+
+| member | content |
+| --- | --- |
+| `png` | the screenshot |
+| `leaf.json` | every visible element: type, role, name and visible text, `rect` and `visible_fragments`, interaction state, parent and reading order, owning application and window |
+| `screentag.txt` | the screen serialized as [ScreenTag](https://github.com/Saidgurbuz/screenparse) markup, on a 0–500 grid normalized to the screenshot |
+| `record.json` | scene metadata (applications, preset, resolution, window stack), eligibility flags, and the action that produced this state |
+
+Two further configurations are available: `transitions` lists every recorded
+click with its before and after observation keys, target element and effect,
+and `preview` is a small stratified sample with images inline (downscaled to
+1024 px) for quick browsing. The field reference is in
+[`docs/schema.md`](docs/schema.md).
 
 ## Loading
 
-Payload is WebDataset; the indexes are Parquet. No custom loader script.
-
 ```python
-import webdataset as wds
-url = "data/train/part-{00000..00903}.tar"
-ds = (wds.WebDataset(url, shardshuffle=True)
-        .decode("pil")
-        .to_tuple("png", "leaf.json", "screentag.txt", "record.json"))
-```
+from datasets import load_dataset
 
-Each observation is four members sharing one key:
+# Stream observations from any split.
+ds = load_dataset("docling-project/DeskForge-1M", split="test_app", streaming=True)
+sample = next(iter(ds))
+sample["png"], sample["leaf.json"], sample["screentag.txt"], sample["record.json"]
 
-```
-<key>.png             the screenshot
-<key>.leaf.json       every visible element
-<key>.screentag.txt   the serialized markup a model is trained to emit
-<key>.record.json     sanitized metadata, and the action that produced this state
+# Transitions: before/after keys, the clicked target and what changed.
+transitions = load_dataset("docling-project/DeskForge-1M", "transitions", split="val")
 ```
 
-The tars are **uncompressed**, so `index/shard_members.parquet` gives a real
-byte offset and length for every member: one seek returns a PNG without reading
-the shard. `examples/` has three runnable loaders — streaming states, pairing
-transitions, and random access by key. `demo/preview.parquet` is a stratified
-sample with images inline for the Dataset Viewer. Its images are **downscaled to
-1024 px on the long edge** so the viewer stays responsive - the canonical
-full-resolution PNGs are in `data/`. ScreenTag coordinates are on a 0-500 grid
-normalized to the viewport, so they still describe the resized image, and each
-row keeps the original `width` and `height`.
-
-Filter before you stream, with Parquet:
+For full passes, read the shards with [`webdataset`](https://github.com/webdataset/webdataset);
+the Parquet indexes select subsets before streaming, and give the byte offset of
+every member for random access:
 
 ```python
+import webdataset as wds
+ds = (wds.WebDataset("data/train/part-{00000..00903}.tar", shardshuffle=True)
+        .decode("pil")
+        .to_tuple("png", "leaf.json", "screentag.txt", "record.json"))
+
 import pyarrow.parquet as pq
 obs = pq.read_table("index/observations/test_id.parquet")
-heavy = obs.filter(obs["occluded_ratio"] > 0.5)   # the degradation slice
+heavily_occluded = obs.filter(obs["occluded_ratio"] > 0.5)
 ```
 
-## How it was made
-
-Xvfb + xfwm4 on Linux, real GTK applications driven headlessly. Annotations come
-from the AT-SPI2 accessibility tree, then are resolved against the window stack
-so a box describes what is *visible*, not what the tree claims. Generation is
-deterministic from a seed; the generating commit and the SHA-256 of both source
-manifests are in `release_provenance.json`, and `checksums/sha256sums.txt`
-covers every published file.
-
-Labels are **automatic**, not human. Quality is enforced by automated audits —
-element coverage against rendered pixels, blank-widget suppression, leaf/window
-ownership, fragment containment — and 59,103 observations that failed them are
-not in this release.
+[`examples/`](examples) contains runnable scripts for streaming states, pairing
+transitions and fetching a single observation by key.
 
-## Actions: what they are and are not
+**Recommended filters.** `state_train_eligible` selects the still-image view
+(excluding near-duplicate scenes and frames where a click changed nothing), and
+`transition_train_eligible` selects the action view, which keeps no-op clicks as
+supervision.
 
-Episodes are **undirected random click explorations**. They are not expert
-demonstrations, not goal-conditioned trajectories, and carry no task
-instruction, reward or success label. The action space is **click only**.
+## Annotation quality
 
-An action is stored on the step it *produced*:
-`observation[k-1] --steps[k].action--> observation[k]`. Pixel coordinates are
-authoritative; `_norm_1000` and `_screentag_500` are derived conveniences that
-round-trip within one grid cell.
+Annotations are generated automatically from the accessibility tree and
+reconciled with the screenshot and the window stack. Captures that fail the
+automated audits (element coverage against rendered pixels, blank-widget
+suppression, window ownership and fragment containment) are not included. A
+human audit finds **99.8%** of sampled element annotations correct.
 
 ## Limitations
 
-One desktop environment (Xfce), synthetic themes imitating Windows/macOS/Linux
-rather than the real thing, click-only actions, no human verification, and
-window-control and window-ownership edge cases documented in
-`docs/known_issues.md`. Because the annotation transcribes what is on screen,
-strings an application printed itself are in it — session paths, synthetic
-persona home directories, the generating host name — all listed with counts in
-`docs/known_issues.md`. No host path appears in any constructed metadata
-record; that is enforced at pack time. Some scenes load live web pages, so a browser window
-shows whatever that site served on the capture date; the domain is recorded.
-Read `docs/known_issues.md` before using this for anything load-bearing.
+All screenshots come from one Linux backend (Xfce): the Windows- and
+macOS-inspired presets reproduce the look of those systems rather than run them.
+Actions are clicks from undirected exploration, not goal-directed
+demonstrations. Because annotations transcribe what is on screen, text drawn by
+applications (for example session paths or live web content in browser scenes)
+is part of the data. Details are in [`docs/known_issues.md`](docs/known_issues.md).
 
-## Licence
+## License
 
-**Not yet determined — see `LICENSE`.** The screenshots contain third-party
-application UI, icon themes, fonts, wallpapers and rendered web pages, and that
-licensing has not been reviewed. Do not redistribute pending that review.
+The dataset is released under the [MIT License](LICENSE). Application
+interfaces, themes, icons, wallpapers and web content visible in the screenshots
+remain the property of their respective owners.
 
 ## Citation
 
 ```bibtex
-@misc{deskforge1m,
-  title  = {DeskForge-1M: dense desktop UI annotation with occlusion-resolved geometry},
-  author = {G\"urb\"uz, Said},
-  year   = {2026},
-  note   = {Version 1.0.0}
+@article{gurbuz2026deskforge,
+  title   = {DeskForge: Dense Supervision from Desktop Environments
+             for Computer-Use Agents},
+  author  = {Gurbuz, A. Said and Nassar, Ahmed and Hong, Sunghwan and
+             Pollefeys, Marc and Staar, Peter W. J.},
+  journal = {arXiv preprint},
+  year    = {2026}
 }
 ```

docs/known_issues.md

@@ -14,17 +14,16 @@ rewriting them would make the annotation disagree with its own image.
 | --- | --- | --- |
 | `/tmp/session-<random>/` | 32.8% of captures | the ephemeral session root of the generating job, in file-chooser paths and window titles |
 | synthetic persona homes, e.g. `/Users/anika/`, `/home/kwame/` | 32.5% | **generated** persona names, not real people |
-| `ess6000-1` | 9.8% | the host name of the generating machine, surfaced as an AT-SPI label |
-| the generating machine's project path | 153 captures (0.01%), 71 shards | `/proj/docling-vision/users/said/desktop_ui/tools/...`, printed by VS Code itself — a terminal showing its own command line, or a docs entry resolving the `code` binary |
-| the dataset author's contact details | **830 captures (0.07%)**, 266 scenes | `Said.Gurbuz@ibm.com;7G5620848;Said Gürbüz`, captured into a document fixture and therefore painted into those screenshots |
+| the generating machine's host name | 9.8% | surfaced as an AT-SPI label in file managers |
+| the generating machine's checkout path | 153 captures (0.01%), 71 shards | printed by VS Code itself — a terminal showing its own command line, or a docs entry resolving the `code` binary |
+| the dataset author's own account details | **830 captures (0.07%)**, 266 scenes | the account name field of the generating user, captured into a document fixture and therefore painted into those screenshots |
 
-The first three were measured over 600 random captures; the author's contact string was counted exactly, across all 1,266,471 source captures. **No host path appears in any of the 1,207,368
+The first three were measured over 600 random captures; the author's account string was counted exactly, across all 1,266,471 source captures. **No host path appears in any of the 1,207,368
 constructed `record.json` files** — that is checked for every record at pack
 time and refused. **No real home directory and no third party's personal data
-appears anywhere.** The project path above appears only where an application
-printed it on screen and the annotation faithfully transcribed it. The author's own contact string is the
-one piece of real personal data, it belongs to the dataset's author, and it is
-disclosed here rather than removed because it is in the pixels.
+appears anywhere.** The checkout path above appears only where an application
+printed it on screen and the annotation faithfully transcribed it. The author's own account string
+belongs to the dataset's author and is disclosed here rather than removed because it is in the pixels.
 
 Downstream filtering on these strings is straightforward: they are exact and
 they are in the text members, not only the images.
@@ -34,13 +33,13 @@ they are in the text members, not only the images.
 Some scenes drive a real Chromium against a real URL, so a browser window shows
 whatever that site served on the capture date. The domain is recorded, the page
 content is not curated, and it may include news text, advertising or images
-belonging to third parties. This is the main reason the licence is unresolved.
+belonging to third parties.
 
 ## Annotation
 
 * **Labels are automatic.** They come from the AT-SPI2 accessibility tree and
-  automated refinement, with no human verification pass. An application that
-  reports its own tree badly is annotated badly.
+  automated refinement; a human audit of sampled elements finds 99.8% of them
+  correct. An application that reports its own tree badly is annotated badly.
 * **Toggle state is invisible on this stack.** AT-SPI 2.40.3 does not expose
   `checkable`/`checked` for menu items here, so a checked menu item is not
   distinguishable from an unchecked one in the annotation.

release_provenance.json

@@ -2,7 +2,6 @@
   "dataset_name": "DeskForge-1M",
   "dataset_version": "1.0.0",
   "created_utc": "2026-08-25T15:15:50Z",
-  "corpus_root": "/proj/docling-vision/users/said/deskshot_corpus/v1",
   "source_manifests": {
     "plan/manifest.jsonl": {
       "sha256": "f50f6a7c87cc6d2a05bb04a820b20e3a634c2cb60ea9ed1afa51325f69c6dc8c",
@@ -14,9 +13,8 @@
     }
   },
   "code": {
-    "repo": "desktop_ui",
-    "commit": "44588e281f2c78f753294592de7c44879a6f89a6",
-    "dirty": false
+    "repository": "https://github.com/Saidgurbuz/deskforge",
+    "release": "v1.0.0"
   },
   "key_scheme": {
     "observation_key": "<shard>__<stem>",