release string | ranking_status string | selection_scope string | deterministic_passed_case_count int64 | deterministic_case_count int64 | p95_turn_latency_ms int64 | automated_eligible_case_count int64 | automated_scored_case_count int64 | relevance_score_out_of_4 float64 | faithfulness_score_out_of_4 float64 | citation_support_score_out_of_4 float64 | automated_judges_are_not_experts bool | not_a_load_test bool | not_an_accuracy_claim bool |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
v0.1 | unranked_single_release | existing_public_drop_versions_only | 195 | 200 | 55,291 | 195 | 195 | 3.5026 | 3.4359 | 2.9897 | true | true | true |
Wiseman Drop Benchmark — v0.1 aggregate results
This public repository contains sanitized, aggregate-only benchmark results. It contains no Drop identifiers, prompts, answers, evidence, credentials, user data, or private source material.
Scope
- Target: Dev EKS application endpoint; fixed concurrency one.
- Cases: 200 internal public-Drop cases.
- This is not a load test, a private-Drop evaluation, or an accuracy claim.
Data contents
This release contains only aggregate benchmark results, methodology, and one safe aggregate JSONL row for the Dataset Viewer. It deliberately contains no row-level evaluation data, corpus text, source documents, prompts, assistant answers, citations, identities, or credentials.
Results
- Deterministic protocol pass state:
False. - Eligible answer-quality cases: 195.
- Automated answer-quality coverage: 195/195.
- Relevance: 3.5026/4.
- Faithfulness: 3.4359/4.
- Citation support: 2.9897/4.
- Initial exact agreement: 43 cases.
- Third-judge adjudication: 152 cases.
Intended use
Use this repository to inspect the disclosed evaluation protocol and aggregate outcomes for the named Dev EKS benchmark run. It is not training data, a retrieval corpus, a public reproduction dataset, or a basis for comparing unreported systems.
Reproducibility artifacts
aggregate-results.jsoncontains sanitized aggregate outcomes and frozen snapshot/manifest fingerprints.METHODOLOGY.mdidentifies the scoring protocol, rubric fingerprint, and judge model families.data/benchmark-results.jsonlsupplies one aggregate-only row for the Dataset Viewer; it is not a corpus, case set, or training split.
License
The aggregate documentation is offered under CC-BY-4.0. This does not grant rights to the private benchmark corpus, source material, or any unreleased row-level evaluation artifacts.
Limitations
The answer-quality scores are produced by automated model judges; no domain expert assessment is claimed. Publication was explicitly authorized by the release owner; it does not represent pending privacy, licensing, reproducibility, or operational-cleanup review gates as complete.
- Downloads last month
- 68