Datasets:
MAGAR-Bench task contract
The public input format is shared across all three task families. Each record
contains a messages array, an images array, and metadata. At inference time
the assistant message is removed. The model must emit one JSON object only,
without Markdown, explanations, wrapper objects, or extra fields.
Assembly Decision
Required output fields:
{
"operation": "place",
"involved_part_ids": ["<part id>"],
"next_subgoal": "place base part"
}
operation is one of align, execute_subtask, inspect, lock, place, or
recover. next_subgoal is one of the canonical values listed in
schemas/schema.json. The high-level task ID is metadata and is not part of the
model output contract.
Atomic Action
Required output is one complete native GUI primitive. The released labels use
click, drag, hotkey, keypress, and type_text. The action type and its
parameters must stay together; future actions and action sequences are not
allowed. Coordinates are screen pixels when the action requires coordinates.
Examples of the native shapes are:
{"type": "click", "button": "left", "screen_point_px": [320, 240]}
{"type": "keypress", "key": "R"}
Minimal Rework
Required output fields:
{
"objective": "minimum_distinct_moved_parts",
"selected_route_id": "<route id>",
"moved_part_ids": ["<part id>"],
"min_moved_part_count": 1,
"action_ids": ["<action id>"],
"full_shortest_path_claim": false
}
The moved-part count must equal the cardinality of moved_part_ids. The
reference is the minimum distinct-part solution inside the validated recovery
candidate set for that state. This benchmark does not claim unrestricted
global optimality over every possible GUI action sequence. Route and action
identifiers are auxiliary recovery information; matching them does not execute
the recovery.
Split and privacy rules
The 20 Train families and six Test families are disjoint. data/test_inputs.jsonl
contains no assistant answers or Test targets. The locked Test gold must stay in
a separate gated evaluator or private artifact.