This dataset consists of frames extracted from third-party video recordings, showing third-party websites and desktop applications. Access is governed by the Restricted Research-Only License and Terms of Access (the LICENSE, full text: https://huggingface.co/datasets/PrentisAI/t2t-v1-annotations/blob/main/LICENSE). If anything below differs from the LICENSE, the LICENSE prevails.

By requesting access you agree to the LICENSE, and in particular that:

you will use the dataset only for Academic Research as defined in the LICENSE: research by an individual affiliated with a university, college, or public or non-profit research institution, intended for publication, and not carried out for or on behalf of any commercial entity;
you will not redistribute the dataset or any derived copy, except as the LICENSE allows (for example, storage on your institution's infrastructure for approved users only, or up to ten images in a publication with personal information redacted);
you will not attempt to identify, locate, contact or profile any person whose personal information appears in the data;
you understand that frames were extracted automatically and may incidentally show personal information; you will not disseminate it, and you are asked to report it to requests@prentis.ai;
any model you release that was trained on the dataset will be released only under terms that prohibit commercial use, and will cite the dataset;
when notified, you will delete removed material, or all copies on termination, within 30 days;
the dataset is provided "as is", without warranty, and the LICENSE is governed by the laws of the State of California.

T2T v1

GUI-agent training data reconstructed from public software tutorial videos by the T2T (Tutorial-to-Trajectory) pipeline. Each video becomes one or more trajectories: the task the video performs, and for every interaction a pre-action frame paired with the action as a tool call. T2T v1 is the training data of T2T-8B, the model of the technical report, released as trained: 345,585 samples in three task views, from 8,838 trajectories of 3,857 videos (desktop, web and mobile).

Access to this dataset is gated. Access is granted immediately after you accept the terms on the request form. Use is governed by the LICENSE (Restricted Research-Only License and Terms of Access): Academic Research only, no redistribution, no re-identification. Please read Data provenance before requesting.

The annotations, system prompts, the training recipe, per-trajectory metadata, and per-image checksums are also available without access approval in PrentisAI/t2t-v1-annotations (no images), so you can inspect the data before requesting access.

Task views

Each view is a separate config.

Config	The model is given	The model predicts	Rows	Size
`navigation` (default)	screenshot, task, previous steps	`<action>` description + tool call	204,445	62.6 GiB
`navigation_stepinstr`	screenshot, task, the current step's instruction, previous steps	`<action>` description + tool call	25,208	7.4 GiB
`grounding`	screenshot, one step instruction	tool call with the target point (or drag start and end)	115,932	35.9 GiB

Examples (assistant turns):

navigation            <action>Click the '+ Create project' button.</action>
                      <tool_call>
                      {"name": "computer_use", "arguments": {"action": "left_click", "coordinate": [578, 123]}}
                      </tool_call>

navigation_stepinstr  <action>Tap the "Tools"/tools menu on the right toolbar.</action>
                      <tool_call>
                      {"name": "mobile_use", "arguments": {"action": "click", "coordinate": [963, 238]}}
                      </tool_call>

grounding             <tool_call>
                      {"name": "computer_use", "arguments": {"action": "left_click", "coordinate": [578, 123]}}
                      </tool_call>

Composition

Rows	Desktop	Web	Mobile	Total
`navigation`	108,406	89,145	6,894	204,445
`navigation_stepinstr`	–	–	25,208	25,208
`grounding`	60,339	50,712	4,881	115,932

Frames: 111,269 desktop, 91,424 web and 25,655 mobile (228,348 in total, 69.5 GiB of JPEG); a frame can appear in several views. Trajectories per source pool: P2 7,719 · P1 724 · P3 317 · P4 78. The pools are those of the technical report: P1 = 720p sources, P2 = 1080p, P3 and P4 = 4K (P4 with a revised extractor). Desktop and web data come from P2 only; navigation_stepinstr adds mobile trajectories from all four pools. Screenshots keep the source resolution (mostly 1920×1080 and 2560×1440, up to 3840×2160). grounding includes 578 drag samples (left_click_drag with start and end).

Schema

Column	Type	Description
`image`	`Image()`	the pre-action frame
`width`, `height`	`int32`	native resolution of the frame
`conversations`	`string`	JSON list of `{"from": "human"/"gpt", "value": ...}`
`task`	`string`	`<view>_<platform>__<pool>`, e.g. `navigation_desktop__mr` (`grounding_drag_...` for drags)
`conv_style`	`string`	conversation template, i.e. which system prompt the sample is meant for
`view`	`string`	`navigation`, `navigation_stepinstr` or `grounding`
`platform`	`string`	`desktop`, `web` or `mobile`
`pool`	`string`	source pool: `parallel` (P1), `mr` (P2), `hr` (P3), `endfix` (P4)
`trajectory_id`	`string`	`<pool>/<video id>[/task_<i>]`; all steps of one trajectory share it
`video_id`	`string`	YouTube ID of the source video
`step`	`int32`	index of the step within its trajectory (0-based)

conversations is stored as a raw JSON string and is not re-interpreted by the loader. It holds the user and assistant turns only: the system prompt of each conv_style is published as plain text in t2t-v1-annotations/prompts/ and must be prepended to reproduce the training setup.

Coordinates are integers normalized to a 1000 × 1000 space, not pixels: x_px = x / 1000 * width, y_px = y / 1000 * height.

Action space

Platform	Tool	Actions in the data
desktop	`computer_use`	`left_click`, `type`, `key`, `scroll`, `left_click_drag`, `wait`, `terminate`, `right_click`, `mouse_move`, `double_click`, `hscroll`, `middle_click`
web	`browser_use`	the above plus `select`, `open_url`, `go_back`, `go_forward`, `triple_click`
mobile	`mobile_use`	`click`, `swipe`, `scroll`, `type`, `wait`, `system_button`, `open`, `long_press`, `key`, `terminate`

Navigation action counts (navigation view): left_click 99,693 · type 29,439 · key 18,972 · scroll 14,150 · left_click_drag 11,435 · wait 8,234 · terminate 7,242 · click 4,453 · right_click 3,232 · mouse_move 2,183 · double_click 1,904 · select 1,481 · other 2,027.

Usage

from datasets import load_dataset

nav = load_dataset("PrentisAI/t2t-v1", "navigation", split="train")      # requires access + token
grd = load_dataset("PrentisAI/t2t-v1", "grounding", split="train", streaming=True)

row = nav[0]
row["image"]                         # PIL image
import json; json.loads(row["conversations"])

Whole trajectories: group rows by trajectory_id and sort by step. Per-trajectory metadata (task instruction) is in t2t-v1-annotations/metadata/trajectories.tsv.

To train with the T2T-8B recipe, unpack the images into the images/ tree that t2t-v1-annotations/meta/*.json points to:

hf download PrentisAI/t2t-v1 --repo-type dataset --include "data/*" --local-dir ./t2t-v1
git clone https://huggingface.co/datasets/PrentisAI/t2t-v1-annotations

import os, pyarrow.parquet as pq, glob
for f in glob.glob("t2t-v1/data/*/*.parquet"):
    t = pq.read_table(f, columns=["image", "pool"]).to_pylist()
    for r in t:
        p = f"t2t-v1-annotations/images/traj_{r['pool']}/{r['image']['path']}"
        if not os.path.exists(p):
            os.makedirs(os.path.dirname(p), exist_ok=True)
            open(p, "wb").write(r["image"]["bytes"])

Data provenance

All frames were extracted from publicly available YouTube tutorial videos of software being operated; video_id and the first directory of each image path are the source video's YouTube ID. The content shown on screen belongs to third parties, not to the maintainers.

Rights in the depicted websites, applications, user interfaces and video recordings remain with their respective owners. The maintainers make no claim of ownership over that content. Prentis AI owns only the annotations and the compilation; the LICENSE licenses those and sets the terms of access for the frames.

Takedown / rights requests: if you are a rights holder and wish content removed, contact the maintainers (see Contact). Removal requests are actioned and a corrected revision published; access may be suspended meanwhile.

Because frames were extracted automatically, some may incidentally show personal information, and some show the presenter's webcam. Please report any personal information you encounter to the address under Contact.

Known limitations

Frames come from recordings, not from a live environment: they can contain the presenter's cursor, webcam overlays, captions, watermarks or a filmed screen, and some user interfaces are not in English.
The grounding view keeps the step's own mouse action (mostly left_click, also right_click, double_click, mouse_move, select, left_click_drag).
The mobile navigation and navigation_stepinstr rows of the P2 pool are the same steps phrased two ways.
Coordinates are normalized to 0–1000 on both axes while resolutions vary; convert with width / height if you need pixels.
No multi-image histories: the history of previous steps is given as text.

Licence

See the LICENSE (the same file is included here and, publicly, in the annotation repository). In short: Academic Research only (as defined there), no commercial use, no redistribution beyond the listed exceptions, no re-identification, and deletion within 30 days of notice. Models trained on the dataset may be released under non-commercial terms. The LICENSE grants no rights in the depicted software, web content or recorded material.

Contact

Access review, personal-information reports and takedown requests: requests@prentis.ai

Citation

T2T is introduced in T2T Technical Report: Synthesizing GUI Agent Trajectories at Scale from Video Tutorials. Citation details will be added on publication.

Downloads last month: 361

Total file size:

114 GB