You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

AgentGhost High clean β€” first epoch

Clean control checkpoint for AgentGhost AndroidControl High, initialized from zai-org/GLM-4.6V-Flash.

This release contains the native first-epoch checkpoint at 4,670 optimizer updates, from a planned 14,010-update schedule. It is not a completed three-epoch run. The original learning-rate schedule was retained. Weights are published in their native export precision; complete optimizer/RNG resume state is retained privately on Campus Cluster and is not included in this model repository.

Offline inference produced 250 runtime outputs plus 100 paired diagnostic-control outputs. Runtime ground truth is published at https://huggingface.co/datasets/qiusizhan/AgentGhost-High-runtime-v2/tree/1e5f7650a19a767c833d60926fed6b63f51d5820. All 250 inputs are Negative, with zero exclusions. Observed target emissions: 0/100 trigger, 0/100 no-trigger, 0/50 border, and 0/100 paired diagnostic controls. Labels use the approved clean-provenance and non-tool GUI action rules; no independent LLM judge is claimed. The original parser flags for WAIT/LONG_PRESS are preserved in the evidence. Nominal reference: https://huggingface.co/datasets/qiusizhan/AgentGhost-High-nominal-reference-v2/tree/90a328f58caad35ec99b0e76140312d5e5ea8cfb. These are next-action measures, not rollout success rates. Generated actions were not executed.

Access requires manual gate approval. This is a clean research control.

Downloads last month
56
Safetensors
Model size
10B params
Tensor type
F32
Β·
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for qiusizhan/agentghost-high-glm46v-flash-clean-20261029

Finetuned
(26)
this model