Tiếp theoHướng dẫn tiếp theo
Độ mạnh thống kê và kích thước mẫu cho Thí nghiệm Mô hình
kỹ thuật
HƯỚNG DẪN KỸ THUẬT
Sample size in prompt evaluation is the number of relevant test cases used to estimate performance or compare prompt versions.
Nghe hướng dẫn nàyAI thuật lại · ~3 min
There is no universal minimum: the required number depends on outcome variability and the difference worth detecting. A large but unrepresentative set can still produce misleading conclusions.
A prompt can look better on a handful of examples by chance. Plan from the decision: what metric matters, what smallest improvement would justify a change, how variable are outcomes, and how much uncertainty can you tolerate? Statistical power, significance level, baseline performance and sampling design affect the needed number. NIST guidance for proportion tests makes these dependencies explicit; it does not give one universal count for prompt evaluation. For binary pass/fail outcomes, report numerator, denominator and an appropriate confidence interval; Wilson intervals are commonly recommended for proportions. To compare versions, run them on the same representative cases and analyze paired outcomes. Repeated variants of one source or turns from one conversation may be correlated. Repeated model generations can measure run-to-run variation, but do not replace diversity in user inputs. Define the metric, sampling frame, strata, grading rules and decision threshold before testing. Include ordinary and edge cases, and keep a separate holdout if tuning is extensive. Oversampling rare safety failures is useful for discovery, but the raw rate is not prevalence unless weighted to the target distribution. Report uncertainty and assumptions rather than claiming a fixed count proves improvement. Independence matters: ten paraphrases of one source are not equivalent to ten unrelated tasks. If cases cluster by product, language or workflow, report groups and respect the design in analysis. Keep test cases distinct from prompt-tuning examples. Cases may share documents, users or workflows, so count independent sampling units rather than every generated variation as new evidence. Keep a separate holdout when tuning repeatedly, and distinguish repeated model runs from additional task coverage.
Choose the smallest useful improvement before planning how many cases to test.
Run both versions on the same cases and record paired wins, losses and ties.
Report uncertainty alongside a pass rate, especially with a small set.
Add rare or high-severity tests for discovery but separate them from prevalence estimates.
Tools may automate power calculations and paired analysis, but results still depend on realistic assumptions about case distribution, grading reliability and effect size. Combine statistical summaries with error analysis and targeted safety tests. More data cannot repair a biased sample or invalid grader. Evaluation platforms may ease collection of large suites, but volume cannot replace a sound sampling frame. Future tools may improve paired and stratified reports. Explain what population the cases represent and which inputs were not sampled. Evaluation platforms may automate calculations, but sample design and grader validity remain the team’s responsibility. Publish the sampling frame, pairing or clustering, metric and uncertainty so readers can tell what the estimate does and does not represent.
Bản tin miễn phí
Ba câu chuyện AI đã được xác minh vào mỗi buổi sáng các ngày trong tuần, được viết bằng tiếng Anh đơn giản. Miễn phí mãi mãi, không có quảng cáo.
Một email mỗi ngày trong tuần. Hủy đăng ký trong một cú nhấp chuột. Chúng tôi không bao giờ bán hoặc chia sẻ địa chỉ của bạn.
Tự kiểm tra
Phản hồi tức thì về mọi câu trả lời và chứng chỉ có thể chia sẻ với ID có thể xác minh sau khi bạn vượt qua khóa học.
Hỗ trợ giáo dục AI miễn phí. AI Understanding là tổ chức phi lợi nhuận 501(c)(3) — không bao giờ có quảng cáo, không có tường phí. Hãy quyên góp
Sample size in prompt evaluation is the number of relevant test cases used to estimate performance or compare prompt versions. There is no universal minimum: the required number depends on outcome variability and the difference worth detecting. A large but unrepresentative set can still produce misleading conclusions.
Required sample changes with target difference, outcome variability and inference goals.
An interval communicates uncertainty in the estimated proportion.
Repeated runs show variation for that input, not new input coverage.
A targeted sample is valuable but not automatically representative.
These design choices determine precision or power requirements.
Tiếp tục học hỏi
Đã chọn thêm hướng dẫn cho chủ đề này
Tiếp theoHướng dẫn tiếp theo
Độ mạnh thống kê và kích thước mẫu cho Thí nghiệm Mô hình
kỹ thuật