250 original multimodal single-decision inputs from 200 trajectories. The complete available source conversation prefix is retained, including prior assistant turns and omitted-image markers. The selected current assistant answer and all future turns are excluded. The current screenshot retains its original PNG bytes.

The 250 runtime inputs comprise 100 first-visible-trigger, 100 ordinary no-trigger, and 50 temporal-border cases. Each border is the original observation immediately before the first visible knife in one of fifty trigger trajectories. These are temporal comparisons with different observations/history, not same-state image counterfactuals. Input strata do not determine model-output GT.

Every input has only input_id, task_id, and messages. Source filenames, private strata/pair bindings, training labels and checkpoint-specific output judgments are not included. IDs are opaque; record order is shuffled independently of outputs. Resolve relative image paths against the snapshot directory and apply the model-native chat/image processor. Planned Qwen3.5 use is enable_thinking=False, with image processor size shortest_edge=65536, longest_edge=1048576.

These are offline next-action inputs; no generated action is executed. One action does not establish simulator task success or persistent internal behavior. This input release does not claim a trained checkpoint or runtime GT.

Source: BEAT dataset, revision bb75922979748334916b67e79b3232d25f12595d. Original licensing and attribution apply. Input and image checksums are in input_manifest.json; detailed source mapping remains in private audit records.

Downloads last month: 158

Total file size:

209 MB