IndicBankBench / VALIDATION.md
Anonymous's picture
Release anonymous IndicBankBench dataset
8392da6
|
Raw History Blame Contribute Delete
3.91 kB
# Candidate release validation
This candidate snapshot was checked with the compatible anonymous harness using only offline
tests and deterministic mock execution.
## Dataset checks
- JSON case files parsed: 799 of 799.
- Unique non-empty `case_id` values: 799.
- Duplicate `case_id` values: 0.
- Required-field and tool-contract lint errors: 0.
- Domains: 6 directory/domain groups (five operational domains plus capability/refusal).
- Primary axes: 20.
- Reference tool definitions: 32.
- Target behaviors: 625 `answer`, 65 `clarify`, and 109 `decline`.
- Largest case file after packaging: 13,069 bytes.
- Files requiring Git LFS under common GitHub/Hugging Face limits: 0.
- Aggregate case-bank SHA-256:
`26fed9a4b801dea83defbdd30637d341a32fe0fbe4c57a39ea10d83fc1813956`.
The aggregate digest is computed by sorting all case JSON files by POSIX relative path and feeding
their raw bytes, without separators, to SHA-256. The value matches `meta.json`.
## Harness checks
- Fresh editable installation under Python 3.9: passed.
- Exact outer runtime/test dependency snapshot: 33 packages in the companion code repository's
`requirements-tested.txt`; it matched the verification environment.
- Dependency consistency and import checks: passed.
- Unit tests against all packaged cases: 51 passed.
- Targeted tool-contract tests: 24 passed.
- Offline scripted smoke case `acct.show_mandates.happy.001`: passed.
- Candidate or judge network calls made during tests: none.
## Linter findings preserved for review
The case linter scanned all 799 cases with 0 errors, 16 warnings, and 115 informational findings.
No benchmark case was changed to suppress these findings.
Warnings:
- 4 dependency-order warnings in `depl.agentic.book_fd.002` and
`depl.agentic.book_rd.001`. The listed gold-call order places a write before declared
prerequisites.
- 9 order-sensitive-list warnings across argument gates.
- 3 weak-affirmation warnings in `depl.create_fd.irrelevant_rag.001`,
`depl.create_fd.irrelevant_rag.002`, and `depl.create_rd.irrelevant_rag.001`; these use the
confirmation-judge fallback.
Informational findings comprise 114 deliberate or reviewable `null` values in visible mock
payloads and one duplicated exposed-tool entry in `cap.social_engineering.009`.
## Authoring-metadata removal
Exactly 8,471 approved underscore-prefixed fields were removed recursively: 7,690 `_comment`,
766 `_variant_note`, 8 `_axis_note`, 2 `_tool_calls_note`, 2 `_formula_note`, 1
`_dependencies_comment`, 1 `_no_write`, and 1 `_variant_note_revised`.
An invariant comparison against each source JSON object confirmed that no other key, scalar value,
object-key order, array length/order, filename, case ID, mock behavior, gold action, grading rule,
or case ordering changed. A second parse confirmed that zero approved authoring fields remain.
The released cases are semantically equivalent to, but not byte-identical with, the source
snapshot.
## Identity and secret scan
All proposed dataset release files were scanned for named or institutional identities,
identity-bearing repository URLs, local absolute paths, private/internal endpoints, private keys,
and high-confidence AWS, GitHub, Hugging Face, OpenAI, Slack, and JWT token formats. No identity,
local-path, private-endpoint, credential, or high-confidence secret hits were found.
Localhost matches elsewhere in the code repository are deliberate unauthenticated examples/tests,
not private services. The synthetic Gmail value in the case bank is benchmark semantics, not a
contributor contact. No dedicated gitleaks, trufflehog, FOSSA, Snyk, or formal legal scan was run.
## License scope
CC BY 4.0 applies to the case bank and authored dataset documentation. No third-party model
responses, hidden reasoning, transcripts, or judge rationales are included in the anonymous data
repository, and no claim is made to relicense such output.