harrrshall/FluidThermal-CUA

flowstate-cua

computer-use data for fluid and thermal simulation analysis.

what this dataset represents

flowstate-cua connects scientific tasks to real desktop screenshots, executed actions, and results that can be checked against the underlying physical fields. it provides 8,260 demonstrations and 269,058 recorded interactions for training and evaluating agents that use scientific software.

the physical fields and task instructions are synthetic. the screenshots and application interactions are real. python controllers operate the native paraview desktop through mouse and keyboard input. scientific tasks also use exported csv files and a bounded calculator. the generation, collection, replay, and scoring code is included.

what it contains

the dataset has two complementary collections:

• scientific tasks: 660 demonstrations and 47,700 interactions across 15 scientific goals and six desktop layouts. tasks ask for quantities and decisions such as overheated volume, outlet mass flow, cooling time, mesh convergence, and design selection. some start from incomplete domains or misleading displays that the agent must investigate.
• foundations: 7,600 demonstrations and 221,358 gui actions over 7,000 two-dimensional physical cases. workflows cover field inspection, contours, calculator expressions, and line profiles.

each collection includes task instructions, screenshots, exact actions and typed contents, physical inputs, saved application states, numerical references, and provenance. scientific tasks additionally includes public tool results and final answers. autonomous model recordings are stored separately from the demonstrations.

a native scientific measurement with a threshold and integration pipeline, numerical output, and an independently checked saved state.

how people can use it

use the demonstrations for next-action training, build an agent that completes scientific desktop workflows, or evaluate whether an agent's analysis supports its final answer.

for scientific tasks, use text and image with the shared system instruction as input, and model_action as the target. the training guide shows how to create a multimodal training example. keep validation and test examples outside training.

start with one example

python -m pip install -U datasets Pillow huggingface_hub

import json
from datasets import load_dataset

data = load_dataset("harrrshall/FluidThermal-CUA", name="v2", split="train", streaming=True)
row = next(iter(data))
print(row["task"])
print(json.loads(row["model_action"]))
row["image"].save("screen.png")

this streams one example and saves its screenshot. use split="validation" or split="test" for the other partitions. name="v2" selects scientific tasks; name="default" selects foundations. these are the supported loader identifiers for the two collections.

the scientific test rows identify their holdout condition in the split field: case, composition, layout, or joint. model_action uses coordinates from 0 to 1000; action retains screen-pixel coordinates for replay. reading examples needs the python packages above. opening saved states and replaying desktop actions also needs the native environment.

download the data

hf download harrrshall/FluidThermal-CUA --repo-type dataset --include "v2/data/*" --local-dir fluidthermal

this downloads the 4.34 gb scientific parquet files. use "data/*" for the 24.51 gb foundations parquet files, or "v2/*" for the 61.22 gb scientific collection, including inputs, recordings, models, source, and reports. omitting --include downloads the complete repository, approximately 155.2 gb. sizes use decimal gigabytes.

release and scale

september 26, 2026 · 8,260 demonstrations · 269,058 recorded interactions

• scientific tasks contains 480 training, 60 validation, and 120 test demonstrations across 138 physical lineages. its 47,700 interactions comprise 45,260 gui actions, 1,780 artifact-tool calls, and 660 final answers.
• its test set covers held-out cases (34), workflow compositions (26), desktop layouts (34), and joint conditions (26). two of the six desktop layouts are reserved for testing.
• foundations contains 6,101 training, 802 validation, and 697 test demonstrations across seven physical families and four workflow types.
• the 648 image-embedded parquet shards total 28.85 gb. complete recordings, scientific inputs, model assets, source, and audit evidence account for the rest of the repository.

the complete release inventory records file sizes and checksums. scientific methods and foundations physics document the generated fields, equations, units, and validation.

model evaluation

the published comparison evaluates the pinned qwen3-vl-4b-instruct model on 120 held-out scientific tasks per setting, using the same tasks, tools, limits, and task-to-worker assignments. it includes 480 autonomous attempts, with failures retained:

• base model: 0/120 tasks completed.
• foundations adapter: 0/120.
• scientific tasks adapter: 20/120 (16.7%).
• scientific tasks adapter with gray screenshots: 0/120.

the base model executed all 21,600 proposed actions with zero rejections. action acceptance and scientific task completion are measured separately. every one of the 20 scientific successes passed independent checks of its answer, native csv exports, saved application state, and source fields.

each adapter trained on 12,000 selected examples for 1,500 optimizer updates; token and image-pixel totals are recorded in the report. the observed scientific-tasks-minus-foundations difference is 16.7 percentage points, with a 95% lineage-cluster bootstrap interval of 8.3 to 25.8 points. these are results for the reported model, task suite, and desktop-plus-artifacts protocol.

the evaluation report provides task-level results, holdout breakdowns, failure accounting, runtime source, and raw recordings. evaluation instructions explain how the checks work.

verification and reproducibility

the release includes independent checks of scientific fields, dataset integrity, split leakage, native outputs, and saved states. foundations also includes 600 numerically checked line-profile exports and 28 successful recorded-action replays. replay results and autonomous-model results are reported separately.

the published streaming example, training example, split loaders, and download commands have been exercised in a clean cloud environment without a hugging face token. all 375 scientific parquet shards and all 776 files in the evaluated recording package passed byte-level checksum checks. the release checks explain the verification scope and link to the evidence.

license and source

mit covers the original code, generated scientific fields, and annotations. see the license and third-party notices. paraview uses a bsd 3-clause license and includes separately attributed components. the qwen-derived adapters and accompanying model assets use apache-2.0, with model license provenance included.

start with the scientific tasks guide, training guide, or foundations training guide. the native environment, portable-input guide, and downloaded replay guide cover reproduction. please cite the dataset url and the repository commit used in your experiment.