PrentisAI/t2t-v1-annotations

T2T v1 Annotations

Annotations, system prompts and metadata for PrentisAI/t2t-v1: GUI-agent data reconstructed from public software tutorial videos by the T2T (Tutorial-to-Trajectory) pipeline. T2T v1 is the training data of T2T-8B, the model of the technical report: 345,585 samples in three task views. This repository contains no images; the frames are in the gated PrentisAI/t2t-v1 repository. It is public so that the data can be inspected before requesting access.

The annotations are licensed for Academic Research only under the LICENSE (Restricted Research-Only License and Terms of Access). You may not redistribute them.

Contents

Path	Contents
anno/<view>/<platform>__<pool>[__drag].jsonl	the samples, one JSON object per line (12 files)
meta/meta_all.json, meta/meta_<view>.json	training recipes: per task the image root, annotation file and conv_style
prompts/<conv_style>.txt	the system prompt of each conversation template (6 files)
metadata/trajectories.tsv	one row per trajectory: video, pool, platform, task instruction, steps per view
metadata/frames.tsv.gz	one row per frame: image path, trajectory, step, platform, resolution, views the frame appears in
stats/image_manifest.tsv.gz	path, size and SHA-256 of all 228,348 frames
stats/task_table.tsv, stats/dataset_stats.json, stats/shard_index.tsv	row counts per task (with the task's name in the original training recipe), totals, the Parquet shards of t2t-v1

Task views

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

228,348 frames from 8,838 trajectories of 3,857 videos: 111,269 desktop, 91,424 web and 25,655 mobile frames. A frame can appear in several views. The mobile navigation and navigation_stepinstr rows of the P2 pool are the same steps phrased two ways (with and without the current step's instruction).

Record schema

{"image": "--LtGxJ8U4E/step_00_start.jpg",
 "width": 1920, "height": 1080,
 "conversations": [{"from": "human", "value": "<image>\nPlease generate the next move ..."},
                   {"from": "gpt",   "value": "<action>...</action>\n<tool_call>\n{...}\n</tool_call>"}]}

• image is relative to the task's root in meta/ (images/traj_<pool>); its first directory is the source video's YouTube ID, and step_NN its step index.
• Coordinates are integers in a 1000 × 1000 normalized space: x_px = x / 1000 * width, y_px = y / 1000 * height.
• Desktop samples call computer_use, web samples browser_use, mobile samples mobile_use; the function signatures are part of the system prompts in prompts/.
• The previous steps of a navigation sample are given as text (Step 1: ...), not as earlier screenshots.

Recipes and prompts

meta/*.json follow the InternVL / Qwen-VL fine-tuning recipe format:

"navigation_desktop__mr": {"root": "images/traj_mr", "annotation": "anno/navigation/desktop__mr.jsonl",
                           "conv_style": "internvl2_5_desktop_navigation_v1", "repeat_time": 1.0, ...}

meta/meta_all.json is the T2T-8B recipe (one epoch over every task, repeat_time 1.0) with the paths rewritten for this release. The conversations hold the user and assistant turns only. To reproduce the training setup, prepend the system prompt prompts/<conv_style>.txt of the sample's task.

Images

hf download PrentisAI/t2t-v1 --repo-type dataset --include "data/*" --local-dir ./t2t-v1   # requires access

Each Parquet row stores the frame's bytes and its annotation path (image.path). To build the images/traj_<pool>/... tree that meta/ points to:

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"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"])

A frame used by several views is written once; stats/image_manifest.tsv.gz gives the SHA-256 of every frame.

Data provenance

All frames were extracted from publicly available YouTube tutorial videos. We hold no rights in the recorded material: rights in the depicted websites, applications, user interfaces and video recordings remain with their owners. Prentis AI owns only the annotations and the compilation. If you are a rights holder and want content removed, write to requests@prentis.ai; we will remove the affected frames and rows and issue a corrected revision. Please also report any personal information you encounter, and do not attempt to re-identify anyone shown.

Citation

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