返回 Skill 列表
extension
分类: 内容与媒体无需 API Key

ml-model-selection

用于在准确性、延迟、成本和可操作性之间进行透明权衡的机器学习模型选择工作流程。当在多个候选模型中选择用于生产环境的模型时使用;不适用于通用API层或仅基础设施的更改。

person作者: jakexiaohubgithub

Ml Model Selection

Overview

Use this skill to choose model candidates with explicit trade-off reasoning, not single-metric optimization.

Scope Boundaries

  • Use this skill when the task matches the trigger condition described in description.
  • Do not use this skill when the primary task falls outside this skill's domain.

Shared References

  • Model selection trade-off rules:
    • references/model-selection-tradeoff-rules.md

Templates And Assets

  • Model comparison matrix:
    • assets/model-comparison-matrix-template.csv

Inputs To Gather

  • Candidate models and benchmark evidence.
  • Serving constraints (latency, throughput, hardware, cost).
  • Risk requirements (robustness, fairness, explainability).
  • Operational ownership and rollback constraints.

Deliverables

  • Candidate comparison matrix with decision rationale.
  • Selected model and fallback candidate.
  • Risk register and rollout recommendation.

Workflow

  1. Capture candidate metrics in assets/model-comparison-matrix-template.csv.
  2. Apply trade-off policy from references/model-selection-tradeoff-rules.md.
  3. Validate decision against production constraints.
  4. Document rejected alternatives and why.
  5. Publish selection and fallback plan.

Quality Standard

  • Selection criteria include accuracy + latency + cost + operability.
  • Decision is evidence-backed and reproducible.
  • Fallback strategy exists for failed rollout.

Failure Conditions

  • Stop when selection ignores production constraints.
  • Stop when alternatives are not evaluated comparably.
  • Escalate when no viable candidate meets minimum requirements.