latent-label-data-augmentation

Latent-label data augmentation repairs benchmarks by rewriting and counterbalancing data to reduce confounds. It fixes latent-label leakage or misalignment that hampers evaluation in mechanistic interpretability tasks.

To reduce confounds in mechanistic interpretability benchmarks, apply structured augmentation moves like rewriting, matched-pair generation, and counterbalancing. This preserves core scenario semantics while improving label validity.

Create matched donor-target pairs by generating controls for causal questions during benchmark augmentation. This preserves core scenario semantics while enforcing documented validation and gap-to-repair traceability.

You need to rewrite benchmark data for latent-label leakage when a benchmark cannot support desired latent labels cleanly. This occurs when misalignment hampers evaluation and requires synthetic augmentation or contrast-set design.

Yes, this benchmark repair approach works for contrast-set design and split construction. It supports mechanistic interpretability tasks by guiding augmentation moves to reduce leakage and improve label validity.

Synthetic augmentation for benchmark repair requires documented validation and gap-to-repair traceability throughout phase-02 workflows. Without enforcing these validation workflows, generated matched pairs may not properly map repairs back to original gaps.