---

name: experimental description: "Routes unstable cleanlab experimental helpers for low-memory batched label finding, span labels, and optional PyTorch examples." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

cleanlab experimental helpers

Use this sub-skill only when the user explicitly wants cleanlab.experimental behavior or accepts an unstable helper because the stable route does not fit their constraint. The cleanlab experimental docs warn that these methods are bleeding edge, may have sharp edges, and are not guaranteed stable across cleanlab versions.

Route here

• cleanlab.experimental.label_issues_batched.find_label_issues_batched or LabelInspector for very large multiclass labels / pred_probs where normal in-memory label-issue detection is too memory hungry.
• cleanlab.experimental.span_classification when the user specifically asks for the experimental span-label wrapper around token classification.
• cleanlab.experimental.mnist_pytorch, cleanlab.experimental.cifar_cnn, or cleanlab.experimental.coteaching when the user wants the optional PyTorch/CIFAR/MNIST/co-teaching examples and accepts extra dependencies, training cost, and instability.
• Opt-in smoke validation with scripts/smoke_experimental.py before relying on these APIs in a fresh environment.

Prefer stable sibling routes first

• Use ../classification/SKILL.md for standard CleanLearning, filter.find_label_issues, count, rank, dataset health, noisy-label benchmarking, and the core low-memory label-inspection idea when the user has not explicitly chosen the experimental batched helper.
• Use ../structured-label-issues/SKILL.md for stable token-classification workflows. Route stable token-label issue detection there instead of using experimental.span_classification.
• Use ../datalab/SKILL.md, ../multiannotator/SKILL.md, ../outlier/SKILL.md, or ../tabular-label-issues/SKILL.md for their stable workflow families.

Read these bundled references

• references/overview.md: task routes, minimal workflows, and the source-artifact bundling decision for this sub-skill.
• references/dependency-matrix.md: required and optional dependencies, including torch, torchvision, and skorch boundaries.
• references/troubleshooting.md: missing optional packages, file-backed array mistakes, slow training/downloads, span wrapper limitations, and experimental deprecation risk.

Operating cautions

• Do not describe these helpers as production-stable defaults or dependency-free.
• Do not make the PyTorch examples part of a stable cleanlab route; they are optional examples, not a required package path.
• For low-memory multiclass inspection, explain the stable classification workflow first, then mention label_issues_batched only when the user needs file-backed or multi-batch processing.
• For span data, state that span_classification is a thin wrapper over cleanlab.token_classification; use stable token-classification guidance for ordinary token-label tasks.