SAD (Situational Awareness Dataset)

Inspect Evals SAD-mini: 2,904 multiple-choice items in five tasks on whether a model knows it is an LLM and can place itself in training, evaluation, or deployment.

Also known as: Situational Awareness Dataset, SAD-mini, Me, Myself, and AI

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorysafety
Subcategorysituational awareness of being an LLM (Inspect Evals SAD-mini multiple-choice slice)
Page statusactive
MetricSAD score (accuracy with invalid answers scored at chance)
Directionhigher_is_better
Dataset size2904
Dataset licenceCC-BY-4.0
PublisherIndependent authors, Constellation, MIT, and Apollo Research; Inspect Evals port by UK AI Security Institute

What it measures

SAD tests whether a language model knows it is a model and can act on that knowledge. The paper's full suite has 7 categories, 16 tasks, and over 13,000 questions (the project site and Inspect README say over 12,000), including self-recognition and tasks that need model-specific facts. This id is the Inspect Evals port of SAD-mini: the five multiple-choice tasks that do not change from model to model. English text. It is not the BIG-bench [self_awareness](self_awareness.md) probes.

Task format

Five Inspect tasks, each a multiple-choice generation with temperature 0. Default shuffle_choices true (gold is A before shuffle). Optional system_variant plain / sp / sp_large and answer_assist prefilling "Answer: (". Lenient prefix match on letter or choice text; invalid format scores random chance 1/n_choices.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub