Compare GUI-agent training datasets

The card distinguishes human task instructions from generated summaries, reasoning, reflections, and quality fields. Generated labels are not independent human quality acceptance records.

• Primary evidence 1

Card maps osworld/train to all_tasks.jsonl and insta/train to task_sequences.jsonl. Six difficulty levels and trajectory screenshots/thoughts/commands are documented. ZIP file labels are not trajectory counts; configurations and composed-task levels are not assumed disjoint.

• Primary evidence 1
• Primary evidence 2

Stage 1 grounding release, Hub revision 5d11f8f5. JSON annotations and ZIP/multipart TAR image packages are publicly listed. The official training configuration groups web, desktop and mobile inputs and selects all rows of its configured inputs; this is a training recipe, not a release sample count or held-out split. The config names omniact.json while the Hub manifest lists omniact_fix.json, so local paths need reconciliation. No sample total is inferred from names such as webui350k. Apache-2.0 is the release card declaration; upstream source rights remain separate.

• Primary evidence 1
• Primary evidence 2
• Primary evidence 3
• Primary evidence 4

Stage 2 trajectory release, Hub revision bd6f1ed4, is distinct from Stage 1 grounding. The public manifest lists source-specific JSON levels and image ZIP/multipart TAR packages. The training recipe uses all rows for some inputs and random:33% for others; several l2 paths occur twice while the manifest also contains l3 files. Recipe rows/levels are neither independent datasets nor disjoint train/test splits. No example/action/trajectory count is inferred. Apache-2.0 is the card declaration, with separate underlying source terms.

• Primary evidence 1
• Primary evidence 2
• Primary evidence 3
• Primary evidence 4

The main repository contains companion training code. The scale and annotation levels here come from the paper, not an independent download or completeness audit.

• Primary evidence 1
• Primary evidence 2

Episodes and action examples are different units. The collection combines multi-step subsets and hindsight-labeled single-step tasks; preserve subset and split choices.

• Primary evidence 1

The manifest lists 58,953 HTML files, which are not converted into transition/example counts. The reported >80k pairs and the HTML-file inventory have different units. The card itself has no sample schema or count; archive/image completeness and JSON contents were not audited.

• Primary evidence 1
• Primary evidence 2

Step instructions can be repeated across multiple actions. Each episode has one more screenshot than action; do not count repeated step instructions as independent task goals.

• Primary evidence 1

The Hub card contains only licensing metadata. The project README supports the episode-training description and rounded count; archive-level quality and exact row counts remain unverified.

• Primary evidence 1

The card identifies deprecated desktop files with screenshot-mapping errors and names corrected releases. Counts are rounded publisher figures for separate platform subsets; instructions and images are not interchangeable.

• Primary evidence 1

The card has license metadata only. The project training README supplies formatting and purpose. No sample count or split inventory is inferred from the training configuration.

• Primary evidence 1
• Primary evidence 2

Rows are not unique screenshots. The card documents image-downscaling conventions that affect coordinates; preserve the device and original image-size fields.

• Primary evidence 1

The card reports a cleaned size of 3.12M without defining the counting unit, so no sample total is inferred. The data card declares noncommercial share-alike terms, separately from the model-weight license.

• Primary evidence 1

The card count is flattened step records, not trajectories or independent tasks. Both successful and unsuccessful trajectories are included according to the project README; no license is declared.

• Primary evidence 1
• Primary evidence 2

Publisher-reported counts cover a selection and reformatting of GroundCUA, not a new disjoint image collection. No independent image-overlap audit was performed.

• Primary evidence 1

A captioning sample is not necessarily a unique video or complete task trajectory. Package layout and split sizes were not established by the inspected sources.

• Primary evidence 1
• Primary evidence 2

Public Hub revision 5a0b35fb includes six bbox JSON files and associated image archives. The training configuration names six input paths with sampling_strategy=all; those input paths are not six disjoint datasets or a verified example count. The card’s source-list wording and mismatched hyperlinks remain unresolved, so its approximately 1M screenshots/10M elements headline is not adopted as a release inventory. No data license is established by the card, and the model repository license is not transferred to these data.

• Primary evidence 1
• Primary evidence 2
• Primary evidence 3

Purpose is established by the official project, not by API tags. No sample records were downloaded; count and data-license fields remain unknown.

• Primary evidence 1
• Primary evidence 2
• Primary evidence 3

The source reports an approximate trajectory total; it does not establish per-platform counts or a complete test-split inventory.

• Primary evidence 1
• Primary evidence 2

The repository gives conflicting episode totals (8,334 and 8,834). This profile uses 8,334 from the specifically recommended current dataset card. No independent archive recount was performed. Coordinates are normalized to 0–1000.

• Primary evidence 1
• Primary evidence 2
• Primary evidence 3

Taxonomy counts describe coverage, not the number of videos or questions. Cases include Android, iOS, desktop software, websites, multi-application use, and XR. Keyframe-selection settings affect evaluation; exact sample inventories are not reported here.

• Primary evidence 1
• Primary evidence 2
• Primary evidence 3

Quantities are rounded publisher descriptions, not downloaded sample counts. Data-license declarations conflict across first-party sources; no permissive license is inferred.

• Primary evidence 1

Quantities are rounded publisher descriptions, not downloaded sample counts. Data-license declarations conflict across first-party sources; no permissive license is inferred.

• Primary evidence 1

Quantities are rounded publisher descriptions, not downloaded sample counts. Data-license declarations conflict across first-party sources; no permissive license is inferred.

• Primary evidence 1

Do not treat Hub size bands or the inconsistent introductory task count as a sample total. Some constituent-domain image packages remain forthcoming.

• Primary evidence 1
• Primary evidence 2

The repository gives conflicting episode totals (8,334 and 8,834). This profile uses 8,334 from the specifically recommended current dataset card. No independent archive recount was performed. Coordinates are normalized to 0–1000.

• Primary evidence 1
• Primary evidence 2
• Primary evidence 3

Coordinates were resized for a specified Qwen2.5-VL image processor and must remain aligned with image resizing. AGUVIS++ material requires images from upstream AGUVIS and OS-Atlas releases; this repository is not a self-contained copy of every referenced image collection.

• Primary evidence 1

Official README calls this a partial training release and says app data will be uploaded gradually. Its training-JSON and test-data links point to separate Hub repositories; those artifacts are not counted as files or splits of xwk123/Mobile3M. Public revision 3df26694 lists app-graph archives; their names do not establish a three-million-example downloadable inventory. Precise inner record/modalities schema is not inspected. The card YAML says Apache-2.0 while its prose and official README assign CC-BY-NC-SA-4.0 to data and Apache to code; the existing data-license value remains unknown with the conflict documented.

• Primary evidence 1
• Primary evidence 2
• Primary evidence 3

Counts describe the card’s first-release claim, not an independently counted archive inventory. The card distinguishes globally image-hash-deduplicated screenshot/VH pairs from the November 2024 complete-app traces; trace state counts must not be added as new unique screens. JPEG/JSON and Parquet are alternative representations. Listed filename ranges are examples, not a complete count proof. No train/validation/test split is documented. Hub revision c734715e is ungated; MIT is its data declaration, without resolving third-party app content rights.

• Primary evidence 1
• Primary evidence 2

Split counts are video-annotation rows; scene and action totals are rounded publisher reports and are not added to row counts. Source videos/images are not included in the Hub files.

• Primary evidence 1
• Primary evidence 2

The raw collection mixes action, language, exploration, and unusable/other traces. These groups have different annotation coverage; not every raw trace is a valid task demonstration.

• Primary evidence 1

Conversation, dialogue-turn and action-step units differ. The released file manifest is inspected; underlying examples were not downloaded.

• Primary evidence 1

The README provides a rounded sample count. It does not establish a unique-screenshot inventory or a normalized split breakdown. Synthetic examples and publicly sourced web content have provenance and quality limits; data terms must be checked separately from repository code.

• Primary evidence 1

These are publisher-reported totals at three distinct grains, not values to sum. Earlier screenshots may be collapsed; each SFT example keeps at most three images. The card and repository disagree on the singular/plural RL filename, so consumers should inspect the actual file manifest.

• Primary evidence 1

These are subset-specific counts, not a total for the full OS-Atlas release. Raw boxes use 0–1 coordinates; the documented training wrapper scales them to 0–1000.

• Primary evidence 1

The repository links distinct AndroidControl-oriented, AndroidWorld-oriented, and web training files plus raw pre-state/action/post-state triples. These are training artifacts, not a documented train/test split, and related representations are not disjoint corpus totals. No numerical inventory is asserted from release names or linked archives.

• Primary evidence 1
• Primary evidence 2

AC and AW refer to distinct target training sets; no sample count is inferred from the two JSONL files or ZIP packaging.

• Primary evidence 1
• Primary evidence 2
• Primary evidence 3

Public manifest exposes data.zip; the internal file schema and number of examples remain unverified. Supervision describes the publisher’s refinement/thought-completion/trajectory-boost pipeline, not an independently sampled data-quality audit.

• Primary evidence 1
• Primary evidence 2

The release contains planner and actor files for two web environments. No sample count is inferred from the package-file manifest. Card loading examples have repository/path inconsistencies.

• Primary evidence 1
• Primary evidence 2

Project-page scope: screenshots paired with augmented view hierarchies, gesture coordinates, layout vectors and metadata. These observations are not action/task totals. The later semantic-annotation download is a related augmentation of screens, not an additional disjoint corpus. The retained original paper reports more than 72k screens from more than 9.7k apps while this project page says more than 66k screens and more than 9.3k apps; the difference is unresolved. Profile counts follow the project page and do not merge these populations. No train/validation/test partition or data license is established here; archives were not downloaded.

• Primary evidence 1
• Primary evidence 2

The inspected README marks the dataset and collection pipelines as in preparation. This profile records an announced research asset, not a verified downloadable corpus; file formats, splits, and release-specific data terms remain unknown.

• Primary evidence 1

The six categories include Web, so they are not six operating systems. Grounding and trajectory annotations follow different collection processes. No sample total or train/test division is inferred from archive-part counts or Hub size tags.

• Primary evidence 1

Observation modes follow publisher archive labels rather than inspected record schemas. No count or training-use permission is inferred.

• Primary evidence 1
• Primary evidence 2

Multiple summaries describe one screen. The repository provides annotations and split files; use the referenced RICO assets for screenshots.

• Primary evidence 1

Raw and Parquet folders are alternative views of the same data. Each sample describes a dragging task over multiple timesteps; action vectors encode button state and coordinates.

• Primary evidence 1

All questions for a given screenshot share one split. ScreenQA Short reuses the same questions/screens; ComplexQA is a separate extension. Do not add these variants as independent original examples.

• Primary evidence 1

The publisher reports these inventories in the companion annotation card. The refusal subset is included in the full recipe. Screenshot access is gated; no image download or independent corpus recount was performed.

• Primary evidence 1

The recipes reuse source images and are not additive datasets. This download contains no screenshots. Counts are publisher-reported; rows named for trajectories are independent decision examples, not full trajectories.

• Primary evidence 1

Counts describe target annotations, not unique screenshots or expanded language/style variants. The evaluation code can reuse a target across configurations.

• Primary evidence 1
• Primary evidence 2

The 8K release name is not an exact record count; the card declares 7,496 examples. Images and original annotations come from OmniAct and are augmented with generated descriptions. Examples are not asserted to be unique screenshots.

• Primary evidence 1

Row count is the card dataset_info value; one row may contain multiple instructions. The card summarizes screenshots as approximately 22K. No data license is declared.

• Primary evidence 1

No sample or episode count is inferred from the size tag or package manifest. Main-branch files were inspected; the card’s v3.0 usage example was not executed.

• Primary evidence 1

Public API revision 33325877 lists Parquet assets but gated=auto; the pinned raw card returned HTTP 401 and no access was requested. Official README identifies this as V1 training data; the API excerpt also announces a Web-Hybrid bounding-box variant with coordinates normalized to [0,999]. That does not establish every shard schema or a release-wide point/bbox count. No train/test partition, example total, or unrestricted downloadability is claimed. Preserve the recorded CC-BY-NC-SA-4.0 data declaration separately from model/code terms.

• Primary evidence 1
• Primary evidence 2

The public manifest contains reward_data.jsonl but no separately packaged image archive. The card declares MIT. The 517k identifier is not converted into a verified sample count, and complete image packaging is not inferred from the JSONL filename.

• Primary evidence 1
• Primary evidence 2

Public JSONL annotations and screenshots are present. MIT is declared in the dataset card. The release name, model-card links and annotation filename use different size labels; this review does not infer a reliable trajectory count from them.

• Primary evidence 1
• Primary evidence 2

The count is explicitly reported by the publisher, not inferred from the filename. Full training data remains outside this release.

• Primary evidence 1
• Primary evidence 2

This is an offline evaluation dataset. Application coverage is not a sample count. The inspected file-layout documentation describes grounding annotations and images; it does not establish a common split or complete action-trajectory inventory.

• Primary evidence 1

Original UI images are retrieved separately by Rico IDs. No release-specific sample count is established.

• Primary evidence 1

Publisher-reported current release, separate from the smaller original-paper population. Screenshot pixels are obtained through referenced RICO IDs.

• Primary evidence 1

The reported video total is distinct from the Dataverse package-file count. The exact-version paper identifies the recorded interface as browser-based Onshape; operating-system compatibility is not established.

• Primary evidence 1
• Primary evidence 2

Card-reported row counts are not unique-screen counts. The source field is important for the per-image license terms; the compilation does not relicense original images.

• Primary evidence 1
• Primary evidence 2

Two purpose-specific files are released: webevolver_policy_sft.jsonl and webevolver_world_model_sft.jsonl. They are not assumed disjoint train/test splits, and no modality or example count is inferred from filenames alone.

• Primary evidence 1
• Primary evidence 2

This is the checklist-generation release rather than the separate preference-pair collection. No unique-task count or data permission is inferred.

• Primary evidence 1
• Primary evidence 2

Preserves the publisher’s split label despite the collection’s training purpose. Rows are not individual preference pairs because rejected is a list. Data license is unknown.

• Primary evidence 1
• Primary evidence 2

Each dataset.csv row is a QA example; web pages and QA examples are different units. Use the v1.0 split table rather than merging the older project headline into it. The card retains email-evaluation instructions, while the manifest now includes test answers; no submission was made. Payload archives were not downloaded.

• Primary evidence 1
• Primary evidence 2
• Primary evidence 3

The OSWorld subset was selected post hoc using model outcomes; no train/test split is supplied. Coordinate calibration, task success, and complete anonymization are not established by package-format checks.

• Primary evidence 1