qiusizhan/AgentGhost-AITZ-nominal-reference-v1

Superseded input protocol

This v1 release does not meet the requested benchmark input contract. It contains 250 trigger-negative references rather than 500 cases from the normal task distribution. Use the corrected v2 inputs. Existing v1 input files and their identities are retained as legacy evidence.

AgentGhost-AITZ nominal reference inputs

250 input-only Android GUI observations with original screenshots and the released AgentGhost prompt format. Native-format screen/action-thought annotation hints are retained. These are offline next-action inputs.

Each record contains only input_id, task_id, and messages; image paths are relative to this repository. No assistant answers, attack labels, poison classes, trigger bindings, model roles, or judge records are included. Model-specific runtime labels are evaluated separately from actual generated responses.

Reference observations are natural, unmodified, released-condition-negative official-test inputs. Their source episodes and screenshot hashes are disjoint from the runtime cohort and training inputs. Selection uses source inputs only, without inspecting model outputs.

Source: Android in the Zoo / CoAT; AgentGhost, pinned commit f6036b09610e8457643d286f057bfefde9c59332. Original data and annotations retain their upstream provenance. Input and image hashes are in input_manifest.json.

Download the gated repository with huggingface_hub.snapshot_download(..., repo_type="dataset"), load inputs.jsonl, resolve each messages[].content[].image relative to the snapshot directory, then use the model's native processor/chat template. For the Qwen3.5 pilot, enable_thinking=False and the original native image processing are used.

Record order is shuffled independently of model outputs. Join separate runtime annotations by input_id, never row position. For the Qwen3.5 AITZ checkpoint, the recorded native image processor size is shortest_edge=65536, longest_edge=1048576 (1M pixels), with the torchvision backend; the model repository stores this same inference setting.