agentnet-partial-and-fail-v1

GUI state transitions (s, a, s') walked on an Ubuntu desktop by Qwen3.8-27B, from tasks taken from AgentNet and run inside OSWorld's Docker environment.

Walks the judge ruled partial or failed. These are the larger half and, for a world model, the more useful one: a walk that did not finish still opened dialogs, switched tabs and changed settings, and each of those is a real transition. Measured over eighty-eight walks, a failure visits 13.5 distinct screens against a success's 7.7.

This repository is the one-condition pool. Tasks are drawn from the 623 Ubuntu-origin AgentNet recordings whose instruction states a single condition, and shards are appended as each batch of a hundred is walked. The 1,050 two-condition tasks will be released separately as v2; a task that failed to walk here rejoins this pool rather than that one.

What a row is

One row is one transition: the screen before an action, the action, and the screen after it.

column	what it holds
state_id / next_state_id	the two states this transition joins
action.executed	the pyautogui that ran, in screen pixels
action_nl	what the agent said it was about to do
action_target	the accessibility element the coordinate landed on -- role, name, bounds. For scoring an answer by element rather than by pixel
transition_role	main for a step along the trajectory, branch for an alternative tried from it. Never summed
candidate_origin	which proposer suggested a branch: uniform-element, policy-sampled, policy-diverse
judged_outcome	success, partial or failure, for the trajectory this row belongs to
split / split_key	train, val or test, assigned 8:1:1 per source recording so a task and every branch of it stay in one split. The same assignment is used by both halves of this release
condition_count / pool	how many conditions the task states, and which pool it was drawn from. v1 is the 623 Ubuntu tasks with one condition; v2 will be the 1,050 with two. Carried on the row so merged shards stay separable
collection_round	which batch of a hundred tasks wrote this row
judged_behaviour	reasonable or unreasonable
hinted / hint_text	whether the agent was given the route a person took, and what it was
walker_version	fingerprint of the code that chose the click, so re-walks sit beside the walks that prompted them

Screenshots and accessibility trees are files beside the rows:

step_000/before.png       step_000/after.png
step_000/before.a11y.xml  step_000/after.a11y.xml
step_000/before.a11y.txt  step_000/after.a11y.txt

Pinning a version

Batches are appended, never rewritten: a shard published in an earlier round keeps its bytes and its filename, and each new batch adds files with the next part numbers. So a copy downloaded last week is still valid, it is simply smaller.

Each batch is tagged, so an exact snapshot can be asked for by name:

from datasets import load_dataset

# the set as it stood after batch one, and it will never change
load_dataset("gui-wm/agentnet-partial-and-fail-v1", revision="round1")

# everything published so far
load_dataset("gui-wm/agentnet-partial-and-fail-v1")

A reader who already holds the whole set can do the same locally by filtering collection_round.

How it was collected

AgentNet task (Ubuntu-origin only)
  -> environment prepared: the file the task names, the package it removes,
     the folder it opens, built or fetched to match what the task asks for
  -> Qwen3.8-27B walks the task, given the accessibility element list and the
     route a person took
  -> at every state, three proposers each suggest five alternatives; each is
     tried from a snapshot and undone
  -> a judge reads every screenshot, what the agent said, what actually ran,
     and the accessibility difference, and rules on the trajectory

Numbers

batches	4
trajectories	230
transitions	57,791
spine (transition_role=main)	4,031
branch	53,760

By application

application	transitions
chrome	8,681
gimp	6,295
libreoffice_calc	6,261
libreoffice_impress	14,160
libreoffice_writer	4,940
os	3,382
vlc	6,463
vs_code	7,609

Branches by proposer

proposer	transitions
policy-diverse	19,270
policy-sampled	14,591
uniform-element	19,899

On the Hub

204 files, 33.6 GB.

Honest limits

• The judge is a model. judged_outcome comes from Qwen3.8-27B reading the walk. Checked by hand on twenty trajectories: about 90% agreed. The two that did not were both cases of the agent being credited for a file the setup had opened for it, which the judge is now told about explicitly.
• Information-lookup tasks are ambiguous. "search for X and tell me about it" has no clear line between having found it and having seen it. Those are in the data and their verdicts are the least reliable.
• Success is not the point of the larger half. A walk that failed can still have visited a dozen distinct screens, and those transitions are as real as any other. That is what agentnet-partial_and_fail-v1 is for.