datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Lite.ScaleCUA
cua-lite/Lite.ScaleCUA
Lite.ScaleCUA grounded teacher trajectories collected on ScaleCUA's OSWorld tasks and judges via the cua-lite lite.scalecua runtime, from two teachers published as separate configs (*.gpt5_5 from gpt-5.5, *.qwen3_8_27b from Qwen/Qwen3.8-27B) and annotated by the same quality pass; ordinary quality gates tagged in metadata.others.exclude_reason, publish-invalid tool leaks/OOB coordinates hard-dropped (filter with not exclude_reason and episode_return>0.5)… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/Lite.ScaleCUA.Lite.CUAGym
cua-lite/Lite.CUAGym
Lite.CUAGym grounded teacher trajectories collected on CUA-Gym task bundles and reward functions via the cua-lite lite.cuagym runtime, from two teachers published as separate configs (*.gpt5_5 from gpt-5.5, *.qwen3_8_27b from Qwen/Qwen3.8-27B) and annotated by the same quality pass; trajectories kept except /opt/env and OOB-coordinate hard-drops, quality gates tagged in metadata.others.exclude_reason (filter with not exclude_reason and episode_return>0.5)… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/Lite.CUAGym.cua-blenderAguvis
cua-lite/Aguvis
cua-lite preprocessed version of Aguvis (xlangai/aguvis-stage1 + xlangai/aguvis-stage2). A large-scale composite GUI dataset using a unified PyAutoGUI action format across mobile / browser / desktop. Stage 1 → grounding.action (single-step locate-and-click across SeeClick, GUIEnv, WebUI, RicoSCA, RICO Icon, Widget Captioning, UI RefExp, OmniACT); Stage 2 → use (multi-step trajectories from AndroidControl, AitW, MiniWoB++, COAT, GUIDE).
Origin… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/Aguvis.easyr1-grounding-dataset-30k-not_grounded-SE-GUI-3B-2MPlite.osworld-assets
assets/synth/ — real-content asset bundle for synth tasks
Real images / audio / video / document templates that synth task pre_config_steps materialize into the container. Hosted on HuggingFace and downloaded at install time (see Lifecycle below); referenced from train.synth.jsonl by stable relative path.
Sibling directories under assets/ may be added later for other bundles (e.g., assets/eval-reuse/ for OSWorld eval-asset copies, assets/shared/ for cross-bundle resources). Each… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/lite.osworld-assets.CUAStepBench
CUAStepBench
CUAStepBench is a benchmark of 278 human-annotated GUI-agent trajectories
for evaluating trajectory-level reward judges. Each trajectory records an
agent attempting a task in a real GUI environment (web, desktop, or mobile),
together with human ground-truth annotations of task success and per-step
quality.
It is the companion evaluation dataset of SeekJudge, a reward framework for
reinforcement learning in computer-use agents.
Homepage… See the full description on the dataset page: https://huggingface.co/datasets/ZJUSCL/CUAStepBench.WebGym
cua-lite/WebGym
WebGym gpt-5.5 trajectories collected on WebGym OmniBoxes tasks and judge via the cua-lite webgym runtime — 25/60/15 site-start difficulty tiers + curated popular-2102 pool, balanced prompt, serp-only+loop cleaned
Origin
https://github.com/microsoft/webgym
https://huggingface.co/datasets/microsoft/webgym_tasks
https://arxiv.org/abs/2601.02439
Load via datasets
from datasets import load_dataset
# entire dataset
ds =… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/WebGym.osworld-trajectoriesCUAGym
cua-lite/CUAGym
CUAGym V2 filtered GPT teacher trajectories for the CUA Lite rollout collection.
Origin
Load via datasets
from datasets import load_dataset
# entire dataset
ds = load_dataset("cua-lite/CUAGym")
# just one platform
ds = load_dataset("cua-lite/CUAGym", "desktop")
# just one (platform, task_type) cohort
ds = load_dataset("cua-lite/CUAGym", "desktop.use")
You can also filter by metadata.platform / metadata.task_type /… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/CUAGym.GUIOdyssey
cua-lite/GUIOdyssey
cua-lite preprocessed version of GUIOdyssey (hflqf88888/GUIOdyssey). A long-horizon cross-app Android mobile dataset of 8,334 task trajectories over ~128k screenshots. Produces two cohorts: use (multi-step agent episodes) and understanding (per-step screen captioning from the source description annotations).
Origin
https://huggingface.co/datasets/hflqf88888/GUIOdyssey
Load via datasets
from datasets import load_dataset
#… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/GUIOdyssey.OpenCUA
cua-lite/OpenCUA
cua-lite preprocessed version of xlangai/AgentNet (OpenCUA). Desktop use trajectories from Ubuntu and Windows/Mac environments with pyautogui-style actions converted to CUA-lite tool calls.
Origin
https://huggingface.co/datasets/xlangai/AgentNet
Load via datasets
from datasets import load_dataset
# entire dataset
ds = load_dataset("cua-lite/OpenCUA")
# just one (platform, task_type) cohort
ds = load_dataset("cua-lite/OpenCUA"… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/OpenCUA.easyr1-103k-4MP-jedi-ui-vision-gta1-data
easyr1-103k-4MP-jedi-ui-vision-gta1-data
Merged dataset composed of the following sources:
datasets/easyr1-63k-nores-jedi-fix-synced-ui-vision-manually-labeled-icon-data-from-yt-4MP (63031 samples in split train)
datasets/easyr1-grounding-gta1-4MP-easy-qwen7b-hard-gta1-7b (39943 samples in split train)
Summary
Generated on: 2025-09-18 06:29:16 UTC
Split: train
Column strategy: intersection
Samples after merge: 102974
Usage
from datasets import… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations-cua-dev/easyr1-103k-4MP-jedi-ui-vision-gta1-data.qwen3-resize-easyr1-110k-bbox0p05-remove-pixmo-uground-seeclickLite.CUAWorld
cua-lite/Lite.CUAWorld
Lite.CUAWorld: agent rollout trajectories staged into the canonical cua-lite layout for SFT distillation.
Origin
Load via datasets
from datasets import load_dataset
# entire dataset
ds = load_dataset("cua-lite/Lite.CUAWorld")
# just one named subset (config)
ds = load_dataset("cua-lite/Lite.CUAWorld", "desktop.use.ardour")
After loading, parse metadata as JSON before filtering by metadata_kind,
dims, or others.*; every row… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/Lite.CUAWorld.CUAGym_V2
Chennzi/CUAGym_V2
CUAGym_V2 raw unfiltered GPT teacher trajectories on remaining CUA-Gym desktop and web tasks. Personal-namespace archive; no quality filtering applied.
Origin
Load via datasets
from datasets import load_dataset
# entire dataset
ds = load_dataset("cua-lite/CUAGym_V2")
# just one platform
ds = load_dataset("cua-lite/CUAGym_V2", "desktop")
# just one (platform, task_type) cohort
ds = load_dataset("cua-lite/CUAGym_V2"… See the full description on the dataset page: https://huggingface.co/datasets/Chennzi/CUAGym_V2.GUI-360
cua-lite/GUI-360
cua-lite preprocessed version of GUI-360 (vyokky/GUI-360), a large-scale dataset of computer-using-agent trajectories on Windows Microsoft-Office apps (Word / Excel / PowerPoint). Provides three desktop task types derived from the successful training trajectories: multi-step use, point grounding (intent -> element coordinate), and screen parsing (listing all interactive UI controls).
Origin
https://huggingface.co/datasets/vyokky/GUI-360… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/GUI-360.mailroom-cuad-contracts
Mailroom Eval: Cuad Contracts
Mirror of the Braintrust evaluation dataset mailroom-cuad-contracts from the
llm-entity-extraction
experiment loop (llm-mailroom legal document pipeline).
Field
Value
Rows
50
Source script
stream_cuad_to_bt.py
Braintrust dataset id
c55ac7f0-56ff-4a2c-b968-7f382ce7daea
Braintrust project id
02fb28b9-60e2-40b6-a68a-b72ee0b237ad
Exported (UTC)
2026-08-22T05:03:53+00:00
Export sha256… See the full description on the dataset page: https://huggingface.co/datasets/Lucius-Morningstar/mailroom-cuad-contracts.CuaGymDesktopGPT55
cua-lite/CuaGymDesktopGPT55
CUA-Gym desktop GPT-5.5 trajectories through batch3 wave12; fixed-task rewards in metadata.
Origin
Load via datasets
from datasets import load_dataset
# entire dataset
ds = load_dataset("cua-lite/CuaGymDesktopGPT55")
# just one (platform, task_type) cohort
ds = load_dataset("cua-lite/CuaGymDesktopGPT55", "desktop.use")
You can also filter by metadata.platform / metadata.task_type /
metadata.others.* after loading;… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/CuaGymDesktopGPT55.easyr1-114k-hard-qwen7b-easy-gta1-4MP-nores-jedi-fix-synced-aug-jitter-tokenizedui-vision-grounding-4MPagentnet-clicks
agentnet-clicks
AgentNet click grounding dataset prepared with agentnet-parse.py.
Generation Details
Script: agentnet-parse.py
Datasets: ubuntu, winmac
Total samples: 109855
Shard size: 1000
Prompt format: gta1
Output format: coordinates
Image resize limit: 4.0 MP
Max actions per trajectory: 5
Sample Entry
Dataset: ubuntu
Prompt:
Click on the GIMP application icon in the left taskbar to launch the image editing software.
Assistant: (35,554)
Bounding… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations-cua-dev/agentnet-clicks.easyr1-110k-bbox0p05-minus-stage3-rl0p2-noise-add-rest-yt-4MP
easyr1-110k-bbox0p05-minus-stage3-rl0p2-noise-add-rest-yt-4MP
Merged dataset composed of the following sources:
datasets/easyr1-103k-bbox0p05-minus-stage3-rl0p2-noise (100155 samples in split train)
mlfoundations-cua-dev/66-yt-app-ui-claude-instructions-no-filter-4MP-gta1-correct-qwen7b-not-correct (7573 samples in split train)
Summary
Generated on: 2025-09-23 15:18:54 UTC
Split: train
Column strategy: intersection
Samples after merge: 107728
Usage
from… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations-cua-dev/easyr1-110k-bbox0p05-minus-stage3-rl0p2-noise-add-rest-yt-4MP.easyr1-grounding-gta1-4MP-easy-qwen7b-hard-gta1-7b
easyr1-grounding-gta1-4MP-easy-qwen7b-hard-gta1-7b
This dataset was generated from filtered GTA shards with images streamed from ZIP archives.
Generated on: 2025-09-18 05:15:30 UTC
Script: push_easyr1_zip_shards_to_hf.py
Filters directory: /p/project1/synthlaion/awadalla1/gta-grounding-data-filters
JSONL glob: gta_shard_*zip.jsonl
Resize max: 4.0 MP
Prompt format: gta1 (output: coordinates)
Random seed: 42
Deduplicate: False
Debug images: True
System Prompt
You are… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations-cua-dev/easyr1-grounding-gta1-4MP-easy-qwen7b-hard-gta1-7b.easyr1-agent-grounding-dataeasyr1-21k-jedi-grounding-4MP
easyr1-21k-jedi-grounding-4MP
This dataset was generated using the EasyR1 grounding dataset pipeline.
Generation Details
Generated on: 2025-08-27 01:12:14 UTC
Script: push_easyr1_to_hf.py
Data directory: /lustre/fsw/portfolios/nvr/users/aawadalla/LLaMA-Factory/data
Parameters Used
Maximum samples: 21000
Image resize (max megapixels): 4.0 MP
Minimum native image resolution: 0.0 MP
Prompt format: gta1_with_resolution
Output format: coordinates
Random seed: 42… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations-cua-dev/easyr1-21k-jedi-grounding-4MP.easyr1-114k-hard-qwen7b-easy-gta1-4MP-nores-jedi-fix-synced-aug-jitter-coord-grid
easyr1-114k-hard-qwen7b-easy-gta1-4MP-nores-jedi-fix-synced-aug-jitter-coord-grid
Augmented version of easyr1-114k-hard-qwen7b-easy-gta1-4MP-nores-jedi-fix-synced-aug-jitter with a fixed 100px coordinate grid overlay.
Each image is overlaid with vertical and horizontal grid lines every 100
pixels at native resolution. Major ticks (every 1 steps)
are emphasized and axis labels show pixel values to help models localize
precise coordinates.
Summary
Generated on:… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations-cua-dev/easyr1-114k-hard-qwen7b-easy-gta1-4MP-nores-jedi-fix-synced-aug-jitter-coord-grid.CUAGym_V1
Chennzi/CUAGym_V1
CUAGym_V1 raw unfiltered GPT teacher trajectories. Personal-namespace archive; no quality filtering applied.
Origin
Load via datasets
from datasets import load_dataset
# entire dataset
ds = load_dataset("cua-lite/CUAGym_V1")
# just one (platform, task_type) cohort
ds = load_dataset("cua-lite/CUAGym_V1", "desktop.use")
You can also filter by metadata.platform / metadata.task_type /
metadata.others.* after loading; every row… See the full description on the dataset page: https://huggingface.co/datasets/Chennzi/CUAGym_V1.CUAVerifierBench
CUAVerifierBench: A Human-Annotated Benchmark for Computer-Using-Agent Verifiers
Universal Verifier paper: The Art of Building Verifiers for Computer Use Agents
Dataset Summary
CUAVerifierBench is an evaluation benchmark for verifiers of computer-using agents (CUAs) — i.e. judges that read an agent's trajectory (screenshots + actions + final answer) and decide whether the task was completed correctly. Where benchmarks like WebTailBench measure agents… See the full description on the dataset page: https://huggingface.co/datasets/microsoft/CUAVerifierBench.cua_debugger_traj
CUA Debugger Trajectories
204 failed computer-use agent (CUA) trajectories on OSWorld, each with a human root-cause annotation.
Three agents were run on OSWorld (Ubuntu desktop, screenshot-only observation, pyautogui execution at 1920×1080). Every trajectory in this dataset is a failure (no task reached evaluator score 1.0). For each trajectory, a human annotator identified the root error step — the earliest step responsible for the failure — and labeled it with an… See the full description on the dataset page: https://huggingface.co/datasets/CyT1ng/cua_debugger_traj.