AgentPROV corpora: Inspect and OpenHands studies (MIT)
收藏资源简介:
Execution-provenance corpora built by AgentPROV v0.1.0 from Inspect logs (ATIA, WhistleBench) and OpenHands trajectories (OpenAgentSafety) published under the MIT licence: each execution as an AgentPROV Execution Record, semantic annotations and W3C PROV bundles, validated and hash-chained. This deposit draft holds 3 AgentPROV corpus version(s), prepared offline by agentprov package --deposit for the release date 2026-10-30. Nothing has been uploaded or published. Corpus atia-benchmark 0.1.0: ATIA: Inspect .eval logs of adversarial tool infiltration through multimodal inputs (mattngyn/atia_benchmark). 110 execution records, of which 0 are outcome-only records. Licence class: MIT. Content digest sha256:cea7d2e1444a96fd8370ef172ec21b60a4a1a83cccc232ce408df4394222521d; root digest sha256:b40dcb3788c8d39bb35fe921a3aa9e5294937ba0b2d636b1140f25f0e405be84. Corpus openagentsafety 0.1.0: OpenAgentSafety: OpenHands trajectories of five models on 352 safety tasks (sani903/OpenAgentSafety). 1558 execution records, of which 0 are outcome-only records. Licence class: MIT. Content digest sha256:9eaa64f1bc715d7091e65eb1059ffb2011112f5cee4cb5da95a5c0a1b3857643; root digest sha256:37d84cd28002e1ad0db59edba736301a48236e433221671a663acdf94f376e52. Corpus whistlebench 0.1.0: WhistleBench: Inspect .eval logs of tool-using agents in corporate-misconduct scenarios (legobridge/whistlebench). 9200 execution records, of which 0 are outcome-only records. Licence class: MIT. Content digest sha256:a5b89089cfad1d75648b9a92268b4835af397033335a18049a30fc3e75ebd235; root digest sha256:11a4c3b8012751a7d4f48661f000ed9af938e819e8934de0e8af647972bfc7d4. Counts are computed from each corpus manifest at sealing. Each DATASHEET states what the records do and do not establish; a statement about the data is scoped as it is there. To verify: compare each archive's sha256 or md5 with DEPOSIT-MANIFEST.json, run agentprov verify on the archives (archive mode, no extraction; H.5.5), and compare each root digest with the one published outside the deposit (the corpus tag annotation and the paper).



