File size: 1,672 Bytes
cd3278e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 | {
"name": "SecJev-4B-Action-Monitor-C1",
"description": "SecJev-4B fine-tuned (variant C1, epoch 1) to judge whether a proposed agent tool call is authorized by the user's request, given the tool outputs the agent has seen (content view).",
"upstream_checkpoint": "SecJev-4B v1.2.0 (UESTC1010/SecJev, GitHub release v1.2.0); upstream records preserved in upstream-provenance.json and upstream-secjev-4b/",
"base_model": "Qwen/Qwen3.5-4B-Base",
"base_revision": "1001bb4d826a52d1f399e183466143f4da7b741b",
"kev_source_commit": "90990a5fac2995b9faa3190f7d437e84f2067768",
"training": {
"code": "ft_action.py in the jev-stageA repository",
"variant": "C1: content view, lr 3e-5, no SecJev-Corpus replay, unauthorized-class weight 3, 1 epoch selected",
"data": "Action-authorization questions built from recorded AgentDojo v1.2.2 runs (build_action_data.py). Train: workspace+travel suites (9,974 questions, 694 unauthorized). Dev: 915. Held-out test: banking+slack (7,588), never used for checkpoint or threshold selection.",
"threshold_reported": "0.35 on P(unauthorized) + 0.5 * P(partially), tuned on workspace+travel normal tasks only"
},
"changed_files_vs_upstream": ["adapter_model.safetensors", "head.pt", "adapter_config.json"],
"hashes": {
"adapter_model.safetensors": "184fbde48a39ca6a8c3f227518581a24e7af2b478db1dac58e028e455284494c",
"head.pt": "93ec77f5536d29dcd468b21ff519e5c2885ddbe0d685ff05f9ab50d1ce892a51"
},
"license": "SecJev Use and Distribution Agreement v1.0 (LICENSE) plus applicable upstream terms (UPSTREAM.md, THIRD_PARTY_NOTICES.md). This is a modified release and is not the original SecJev-4B."
}
|