[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"glossary-stratified-sampling::en":3,"gloss-cluster-stratified-sampling::en":23,"gloss-next-stratified-sampling::en":9},{"slug":4,"category":5,"name":6,"definition":7,"meta_desc":8,"faq":9,"schema_markup":9,"related":10},"stratified-sampling","mlops","Stratified Sampling","Stratified sampling means dividing the population into groups that matter and drawing from each deliberately, instead of taking cases at random and hoping the mix comes out useful. For evaluation sets it is usually the right default, because a purely proportional sample is dominated by whatever is common, and what is common is rarely what you are worried about. A support classifier will be measured almost entirely on routine questions if the set mirrors traffic, while the rare categories where a mistake is expensive appear a handful of times and their scores are too unstable to read. The strata are chosen from what you would want to make a decision about separately: customer segment, language, document type, request length, difficulty, the categories in your failure taxonomy, whichever of those a stakeholder would ask about. Then you pull enough of each to say something about it, which usually means over-representing the rare ones on purpose. The trap comes at reporting time. A stratified set is no longer a picture of traffic, so an overall average computed across it describes a population that does not exist. If you want a figure that represents live behaviour, weight each stratum back to its real share when combining; if you want to know how the system handles the hard cases, report those separately and say so. The mistake that keeps recurring is building the set carefully, over-sampling the difficult categories for good reasons, and then quoting the unweighted mean in a summary, where it reads as a general quality number and understates the system considerably. Record the strata and their sizes alongside the set, not in someone's head. Six months later the numbers will be compared with new numbers by people who were not there, and a set whose composition is undocumented invites exactly the misreading described above.","Stratified sampling draws from groups that matter instead of mirroring traffic. Why evaluation sets need it and why the unweighted average then misleads.",null,[11,14,17,20],{"slug":12,"name":13},"class-imbalance","Class Imbalance",{"slug":15,"name":16},"data-sampling","Data Sampling",{"slug":18,"name":19},"golden-dataset","Golden Dataset",{"slug":21,"name":22},"survivorship-bias","Survivorship Bias",[24,28,31,35,38,41,44,47,50,53,54,57],{"slug":25,"category":5,"name":26,"updated_at":27},"acceptance-threshold","Acceptance Threshold","2026-09-10T03:30:02+00:00",{"slug":29,"category":5,"name":30,"updated_at":27},"adjudication","Adjudication",{"slug":32,"category":5,"name":33,"updated_at":34},"annotation-guidelines","Annotation Guidelines","2026-09-08T21:32:36+00:00",{"slug":36,"category":5,"name":37,"updated_at":34},"baseline-model","Baseline Model",{"slug":39,"category":5,"name":40,"updated_at":34},"batch-inference","Batch Inference",{"slug":42,"category":5,"name":43,"updated_at":27},"blind-evaluation","Blind Evaluation",{"slug":45,"category":5,"name":46,"updated_at":34},"canary-prompt","Canary Prompt",{"slug":48,"category":5,"name":49,"updated_at":27},"ceiling-effect","Ceiling Effect",{"slug":51,"category":5,"name":52,"updated_at":34},"champion-challenger","Champion-Challenger (A\u002FB Model Testing)",{"slug":12,"category":5,"name":13,"updated_at":34},{"slug":55,"category":5,"name":56,"updated_at":34},"continuous-batching","Continuous Batching",{"slug":58,"category":5,"name":59,"updated_at":27},"coverage-gap","Coverage Gap"]