500 input-only AndroidControl observations with original PNG bytes, device status, chronological action history and the released prompt format. Low retains the per-step instruction; High uses the high-level goal.

500 original, unmodified official-test cases are sampled uniformly without replacement. The sampling unit is an episode/step case (403 distinct episodes); multiple steps from one episode remain grouped by task_id. There is no trigger, source-target or model-output filter. Naturally occurring trigger conditions are retained. Training and evaluation episodes/images are excluded; declared source coordinate/geometry quality filters include prior history. This is a normal-task-distribution reference, not a guarantee that every checkpoint produces benign output.

Only input_id, task_id and user messages are detector-visible. No assistant answers, trigger bindings, model roles or runtime GT are published here. Join separately generated checkpoint-specific labels by input_id. These are offline next-action inputs; no actions or emulator commands are executed. Selection is fixed before model generation.

Runtime and reference episodes and screenshot hashes are disjoint. Resolve image paths relative to the downloaded snapshot and use the checkpoint-native processor. Planned Qwen3.5 encoding uses enable_thinking=False and shortest_edge=65536, longest_edge=1048576. Training, native tensor-cache checks and actual-output labels remain separate stages.

Sources: Google Research AndroidControl and AgentGhost. Immutable source identity and input/image hashes are recorded in input_manifest.json. V1 is a historical noncompliant pilot and is not the benchmark input protocol.

Downloads last month: 198

Total file size:

336 MB