puyang2025's picture
docs: clarify dataset metadata scope
5cbb9f0
|
Raw
History Blame Contribute Delete
3.14 kB
metadata
pretty_name: Agent Authorization Dataset
language:
  - en
license: other

Agent Authorization Dataset

Versioned evaluation and research data for the Agent_Action_Security project. The current release contains deterministic authorization cases for OpenAI-compatible function-tool traffic. It is used for regression, conformance, generated-corpus, benchmark, and optional planner robustness evaluation; it is not runtime policy or production audit data.

Files

File Purpose Records
data/authz_cases.yaml Hand-authored deterministic AuthZ regression and demo cases 54
data/authz_conformance_golden_v1.yaml Reviewed authz_dataset_v1 conformance matrix 10
data/authz_generated_cases.jsonl Reproducible synthetic authz_test_case_v1 corpus 400
data/real_llm_prompts.yaml Optional real-LLM planner robustness prompts 14 scenarios

The authoritative schemas and validators remain versioned with the application code:

Provenance and reproducibility

These files were imported without content changes from AnshiHuawei/Agent_Action_Security@19fb18f9e557771de4488e06b9f6b86d2e4ef561. Their import checksums are:

File SHA-256
authz_cases.yaml 35e0981d7b367647a8bb5fc0c2c4e1b51b803d9e5ffdc2f99f9b0434786a446d
authz_conformance_golden_v1.yaml e36cc253fd7a54229d632954c250a9e8e029550a698bd78a1a0052f189d9d50b
authz_generated_cases.jsonl 140f81f72ef399c82da8a6e9e4fd35fefcb00e3d8a0fb8851d54909c41346756
real_llm_prompts.yaml 5d4dab45bd3155416838e83210645ee8bb4ce040a9fbd515fe06729326cd6ce6

authz_generated_cases.jsonl is regenerated deterministically with seed 20260713 and generation version authz-corpus-v1 by scripts/generate_authz_cases.py in the application repository. Each authz_dataset_v1 conformance record contains its own source, transformation, human-review, policy-version, and license-review metadata. The synthetic authz_test_case_v1 records instead carry deterministic generator, seed, split, generation-version, and policy-version metadata.

Privacy and security

The corpus uses fictional demo identities and synthetic authorization scenarios. It must not contain production prompts, access tokens, secrets, real salary records, or personally identifiable information. Runtime Decision Log entries are a separate data class and must not be copied here without the project's reviewed sanitization and dataset-admission process.

License

The dataset is Project-Proprietary. Use, redistribution, or derivative publication requires permission from the dataset owner. Upstream-derived records additionally retain their per-record provenance and license-review metadata.