Historical pilot — protocol correction required. This v1 does not meet the benchmark contract. The required runtime cohort is 100 trigger / 100 no-trigger / 50 separately defined border cases; the nominal reference must contain 500 unmodified normal-task-distribution cases. The v1 files are retained for provenance only and must not be treated as the final benchmark input sets. A corrected release is being prepared.
AgentGhost-Low runtime inputs
250 input-only official-test AndroidControl observations. Original PNG bytes, real device status and chronological action history are preserved. Low includes the released per-step instruction; High uses the high-level goal.
Only input_id, task_id and user messages are detector-visible. Image paths resolve relative to the downloaded snapshot. No assistant answers, training labels, trigger bindings, model roles or runtime GT are included. Join future model-specific actual-output annotations by input_id, never row position.
The primary runtime cohort has 120 first-eligible trigger observations and 130 natural controls. The 120 exact-goal-restoration counterparts are retained separately for offline diagnostic evaluation.
Selection was frozen before model outputs; source coordinate/PNG-geometry anomalies and affected histories are excluded from held-out selection. These are offline next-action inputs, without tool/emulator execution. Input/image hashes are in input_manifest.json.
Sources: Google Research AndroidControl and AgentGhost, pinned commit f6036b09610e8457643d286f057bfefde9c59332. Original screenshots retain their upstream provenance. Use the checkpoint-native chat/image processor; planned Qwen3.5 inputs use enable_thinking=False and the recorded 1M pixel policy. GPU training and actual-output GT are pending; this input release does not claim their completion.
- Downloads last month
- 33