contra-labs/gemini-creative-campaign-trajectories

Gemini Creative-Campaign Computer-Use Trajectories (Preview)

A preview release of computer-use agent trajectories from creative-campaign work in Google Gemini (generative image creation and editing, in Chrome), with some segments crossing into Adobe Photoshop and Notion. Each operator, a designer sourced from Contra's top-earning talent works a brand brief into a sequence of deliverables across multi-segment sessions: a hero image, secondary assets, local edits, a download, and a final submission upload. Every step pairs a screenshot with a structured action, a tool_call, an executable code string, and reasoning augmented from the operator's spoken narration. Every trajectory also carries an expert quality evaluation (alignment, efficiency, difficulty, and a written rationale).

A sample of the human data Contra Labs builds on demand. We are an independent human-data and creative-evaluation lab: expert evaluation, rankings, and benchmarks for AI outputs, plus custom datasets like this one. See Working with Contra Labs.

Motivation

Most open computer-use and GUI-agent datasets focus on short, single-goal web or OS tasks. This preview captures long-horizon generative-media work. The operator iterates a brand brief across many prompt-and-generate cycles in Gemini, switches between apps to finish and submit assets, and judges each deliverable against the brief. Alongside the trajectory, each episode carries an expert evaluation of how well the workflow met its goal.

What a session contains

A session follows one operator building a campaign end to end. The campaigns span consumer verticals, including a grill brand (TRENTI), an outdoor-gear brand (THE URBAN NOMAD), activewear (Sola), a snack brand (Farmly), fashion (MUXO), a furniture brand, and the original sun-care (SOLENE) and ice-cream (Bare Berry Rush) campaigns. A single session is split into segments, each a self-contained sub-task: generate and refine the hero image, create a secondary asset, apply a local edit (such as swapping a product or removing an object), and download or finish the selected result. Some sessions cross into Photoshop for finishing or Notion for submission. Group by session_uuid to recover a session, and by trajectory_uuid (ordered by step_index) to recover one segment-trajectory.

Format and schema

The dataset is a single Parquet file with one row per step: 266 steps across 28 trajectories from 11 sessions (3 to 40 steps per trajectory; sessions of 1 to 5 segments), recorded on macOS and Windows. Screenshots are embedded in the Parquet as a Hugging Face Image feature. Trajectory-level fields are repeated on every step so a row is self-contained. Each trajectory ends with a terminate step.

Each step's image is the screenshot the operator saw when taking that step's action — the pre-action observation, not the post-action result. So a step's code acts on the state in its own image; the result of that action appears in the next step's image, and that step's reflection describes it. This screenshot-then-action pairing is the standard for computer-use corpora (the model learns observation -> action). For example, a "remove the banner" step shows the image that still has the banner; the banner-free render is the next step's image.

from datasets import load_dataset

ds = load_dataset("contra-labs/gemini-creative-campaign-trajectories", split="train")
row = ds[0]
print(row["instruction"])
print(row["action"], "->", row["code"])
row["image"].show()

Per-step fields:

Column	Description
trajectory_uuid	stable UUID for the trajectory (one segment; group on this to recover a segment-trajectory)
session_uuid	stable UUID for the recording session, shared by every segment-trajectory from that session
step_index	position of the step within its trajectory (contiguous, 0-based)
image	the step's screenshot (Hugging Face Image feature)
thought	the step's reasoning
action, action_type	human-readable action and its taxonomy label (write, click, press, hotkey, rightClick, navigate, wait, terminate)
code	executable pyautogui-style action string
tool_call	structured action intent as a JSON string, { "name", "arguments" } (null on terminate)
execution_paths	deployment hints as a JSON string (keyboard_shortcut / mcp_tool / menu_path; all three are null throughout this browser-only sample)
source	application surface the step occurred in (browser; null on the terminate step)
coord_x, coord_y	normalized pointer coordinates when applicable, otherwise null
observation, reflection	on-screen state before the action, and an assessment of the resulting state
viewport_width, viewport_height	screen dimensions for the step
timestamp_ms	step time offset

Trajectory-level fields (denormalized, repeated on every step):

Column	Description
instruction, natural_language_task, actual_task	the task stated at increasing levels of specificity
reason	the evaluator's written rationale for the scores
task_completed	whether the deliverable and evaluation were completed
alignment_score, efficiency_score, task_difficulty	expert evaluation scores (alignment 5-8, efficiency 4-8, difficulty 2-7 in this sample)
domain, platform, task_category	Professional; macOS or Windows; compose_generative or photo_organize
applications_used, duration_ms, step_count	session summary

The fields come from three sources. thought, observation, and reflection are model-augmented from the operator's narration transcripts (the operator narrated each step aloud as they worked). action, code, tool_call, coordinates, and the screenshots are the operator's real execution. alignment_score, efficiency_score, task_difficulty, and reason are the expert evaluation.

Relation to computer-use research

Each step follows the per-step pattern shared by computer-use corpora such as AgentNet (and related work like AgentTrek): a screenshot, a first-person observation and thought, a structured action (an action-type label, a pyautogui-style code, and a tool_call), and a post-step reflection. Using a familiar shape keeps the data interoperable with existing computer-use tooling. The additions specific to this set are generative-media tasks in Gemini, multi-segment session grouping (session_uuid / trajectory_uuid), and a per-trajectory expert evaluation.

Curation

Each trajectory was cleaned to a single coherent path: low-value steps (misclicks, no-ops, abandoned exploratory branches) were dropped and indices renumbered contiguously, and every trajectory ends with a clean terminate step. Each step was reviewed against its screenshot so the action, observation, and reflection match what is on screen, with click coordinates checked against the controls they target and terminal claims kept to what the captured frames show; trajectory outcomes were verified against the frames and corrected where a raw label was wrong. Typed prompts were verified against the session's DOM recording (rrweb) so the recorded text matches what the operator entered, preserving original spelling. Step counts and per-step fields are kept consistent with the screenshots.

Intended use

• Exploring long-horizon, multi-segment generative-media trajectories as training and evaluation data.
• Studying creative computer-use workflows (prompt-iterate-evaluate loops) and how expert evaluation scores relate to the trajectory.
• Previewing the trajectory + evaluation schema, with tool_call grounding and session structure, for creative-professional work.

Limitations and scope

This is a small preview corpus of 28 trajectories and 266 steps from 11 sessions in a single vertical (generative creative campaigns in Gemini) on macOS and Windows. The reasoning fields are model-augmented from the operator's narration transcripts rather than verbatim transcripts. It suits evaluation and qualitative study more than large-scale training.

Working with Contra Labs

Contra Labs is an independent human-data and creative-evaluation lab, backed by a network of verified creative and domain experts. This dataset is one example of our work.

We partner with AI teams on:

• Evaluation, rankings, and benchmarks. Expert human judgment on model outputs across text, image, video, audio, UI, and multi-modal work, scored for quality, style, and brand fit.
• Custom dataset creation. Computer-use trajectories, pairwise preference data, and evaluation scores, custom built to your domain, schema, and difficulty.

To commission an evaluation or dataset for your domain, reach out through contralabs.com or email partnerships@contralabs.com.

Provenance and consent

Recorded with consent from the participants. Screenshots were reviewed and the operator's own personally identifying information - account avatars, names and greetings, personal names in private app chrome, page-footer org names, IP-derived location lines, and device or drive labels - was redacted with an opaque marker. Stock, AI-generated, and third-party imagery was left intact, as it does not identify the operator: the AI-generated campaign deliverables on the canvas, reference and stock photos, sample thumbnails, and the brand, product, and tool names shown are all part of the work itself.

References

• AgentNet computer-use dataset: huggingface.co/datasets/xlangai/AgentNet. OpenCUA: Open Foundations for Computer-Use Agents, arXiv:2508.09123.
• AgentTrek: Agent Trajectory Synthesis via Guiding Replay with Web Tutorials. ICLR 2025. arXiv:2412.09605.

Citation

@misc{contra_labs_gemini_creative_campaign_trajectories_2026,
  title        = {Gemini Creative-Campaign Computer-Use Trajectories (Preview)},
  author       = {Contra Labs},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  note         = {Preview release}
}

License

Released under CC-BY-4.0. Free to use with attribution to Contra Labs.