Unverified50% confidenceSolutionExact time
Stratified evaluation divides the evaluation dataset into multiple slices and sets independent pass thresholds for each slice, making it possible to capture significant degradation in subgroups.
1
Sources
50%
Confidence
Long-term
Relevance
8/8/2026
First Seen
Sources
从生产事故到评估用例：LLM回归测试的实战方法论
Related Claims
UnverifiedStratified evaluation calculates metrics separately by scenario type, user group, or business risk level, and can capture significant degradation in subgroups.80% similarUnverified等价类划分将输入数据按有效/无效归类，从每个等价类中各取一个代表性数据进行测试，以在有限用例数量下最大化覆盖率74% similarUnverifiedStratified sampling draws samples according to the proportion of each class in the population, ensuring that the class distribution of the subset is consistent with the original dataset73% similarUnverifiedEvaluation platforms such as Braintrust and Humanloop have built-in slice-based evaluation functionality, allowing teams to see both the overall score and the independent performance of each slice in a single evaluation run.66% similarUnverifiedThe model evaluation feature supports building custom datasets, automatic scoring, multi-dimensional metric measurement, and customizable scoring criteria.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/709148API
curl https://kongchang.com/api/v1/knowledge/claims/709148MCP
get_claim(id=709148)