Add dataset card: AgentNet task source, 81 trajectories, policy branches

README.md

@@ -0,0 +1,119 @@
+---
+license: apache-2.0
+task_categories:
+  - other
+language:
+  - en
+tags:
+  - gui
+  - world-model
+  - desktop
+  - agent
+  - ubuntu
+  - branching
+size_categories:
+  - 10K<n<100K
+configs:
+  - config_name: default
+    data_files: data/*.parquet
+---
+
+# GUI World Model — Branch Transitions
+
+`(s, a, s')` transitions taken **beside** a walked trajectory rather than along
+it: the same state, a different action, and the screen that actually followed.
+Every state is captured from a running Ubuntu desktop — a screenshot, the
+accessibility tree as XML, and the rendered element table the model reads.
+
+A spine records what an agent did. This set records what it did **not** do and
+what would have happened, which is the question an agent asks a world model
+when it plans.
+
+## What is here
+
+| | |
+|---|---:|
+| transitions | 57,810 |
+| trajectories branched | 81 |
+| shards | 341 |
+| size | ~32 GB |
+
+**All 81 trajectories carry AgentNet instructions**, nine from each of nine
+applications:
+
+    chrome  gimp  libreoffice_calc  libreoffice_impress  libreoffice_writer
+    os      thunderbird  vlc  vs_code
+
+Nine each is deliberate — the collector takes one task from each application in
+turn, so a run that stops early still covers every application rather than the
+beginning of the alphabet.
+
+Several tasks appear more than once under different start states
+(`__cfg0` … `__cfg3`), which is how the same instruction is walked on more than
+one arrangement of the desktop.
+
+## Spine and branch are both present, and are not the same thing
+
+Sampled over 250 rows spread through the set:
+
+| `origin` | share |
+|---|---:|
+| `branch` | 90% |
+| `spine` | 10% |
+
+The spine rows are the walked path the branches hang off. **Do not add the two
+counts into a single figure** — they answer different questions, and a total
+that mixes them cannot be attributed to either.
+
+Every branch in this set is `candidate_origin: policy-sampled` at
+`branch_depth: 1`: one step away from a spine state, with the action drawn from
+the policy rather than uniformly from the clickable elements. The uniform arm
+lives elsewhere.
+
+## Columns that say what a row is
+
+| column | values |
+|---|---|
+| `origin` | `spine` · `branch` |
+| `candidate_origin` | `policy-sampled` (branch rows) · empty on spine rows |
+| `branch_depth` | `1` on branch rows, empty on spine rows |
+| `trajectory_source` | `replay` throughout |
+| `source_archive` | `walked-spine` — the branches were taken from stored spine states |
+| `task_id` | `agentnet_<application>_<id>[__cfg<n>]` |
+
+### A known gap
+
+`instruction_source` is filled on spine rows and **empty on branch rows**. The
+task's origin is still recoverable from the `task_id` prefix, which is
+`agentnet_` for every row here, but a reader filtering on the column alone will
+find the branches unlabelled.
+
+## What this set does not tell you
+
+**Whether the branched trajectories succeeded at their tasks.** These were
+collected before `spine` was narrowed to mean *a trajectory a judge called
+successful*, so a branch here is an alternative to an action that was taken,
+not necessarily an alternative to a correct one.
+
+That matters for evaluation and not for training. Training on a transition uses
+only the fact that this screen, under this action, became that screen — which is
+observed, and true regardless of how the task ended. Evaluation that asks
+whether a branch was better or worse than the path needs the path to be known
+good, and here it is not recorded.
+
+## How the branch actions were chosen
+
+The policy is asked once for the state with `n=32` at temperature 1.0, and the
+completions are grouped by **intention** rather than by string: the verb plus
+the smallest accessibility element containing the point. On one Thunderbird
+state, 24 distinct strings were three intentions, six of them the same button
+five pixels apart. The most common intentions are executed, up to a budget of
+ten, because each execution costs a VM action and a snapshot.
+
+`draw_count` — how many of the 32 produced each intention — is computed but not
+stored, so a nine-of-thirty-two action cannot be told from a one-of-thirty-two
+one after the fact.
+
+## Related
+
+- `gui-wm/spine` — 154,220 spine-only transitions, no branches