HHH Evaluation Set (binary comparisons)
Constitutional AI: Harmlessness from AI Feedback — refined
A benchmark of binary response-comparison questions testing whether a model can identify the more helpful, honest, and harmless reply; originally 221 comparisons from Askell et al. (2021) on BIG-Bench, extended here with 217 additional harder harmlessness-focused comparisons.
Saying a model understands helpful, honest, and harmless behavior is easy; testing it is harder. This benchmark does the testing, showing a model pairs of candidate replies and asking it to pick the better one, with most questions inherited from earlier work and a harder harmlessness-focused set added here.
What the benchmark tests and where its questions come from lead into what it is used to check, connecting to the HHH framework it operationalizes and the BIG-Bench suite that first hosted it.
What it tests, and where it comes from
The HHH evaluation set is a benchmark of binary response-comparison questions: given a conversation and two candidate assistant replies, the task is to pick the one that is more helpful, honest, and harmless. It operationalizes the qualitative Helpful, honest, harmless (HHH) alignment framework into something a model's accuracy can be scored against, turning a definition into an instrument (Helpful, honest, harmless (HHH) alignment framework — is operationalized as → HHH Evaluation Set (binary comparisons)). Its original 221 comparisons come from Askell et al. (2021) and are hosted publicly as a subset of BIG-Bench (HHH Evaluation Set (binary comparisons) — is hosted as a subset of → BIG-Bench, constitutional-ai, §"B Identifying and Classifying Harmful Conversations", p. 20); Constitutional AI extends this with 217 new, deliberately harder comparisons, because models were already scoring "well over 90% binary accuracy" on the original set, with the new items "primarily focusing on more subtle tests of harmlessness, including examples where an evasive response is disfavored over a harmless and helpful message" (constitutional-ai, §"2 Evaluating the Potential for AI Supervision of HHH", p. 6). The extension is distributed only through the paper's own repository rather than folded into BIG-Bench itself, so "the HHH eval" as Constitutional AI uses it is really a public core plus a private harder addendum.
What it is used to check
The benchmark does double duty in the paper. It is where Chain-of-Thought (CoT) Prompting is shown to significantly improve a model's ability to pick the better response (Figure 4), and it is where Model Calibration (well-calibrated multiple-choice probabilities) moves from an assumed property to a demonstrated one: "we show calibration of the RL-CAI labels...on our new HHH eval. We find that the feedback model's log-probabilities are reasonably well-calibrated" (constitutional-ai, §"4.3 Main Results", p. 12) (Model Calibration (well-calibrated multiple-choice probabilities) — is empirically verified on → HHH Evaluation Set (binary comparisons)). Both uses sit inside the Helpful, honest, harmless: a shared definition of alignment theme, while the benchmark's own two-stage life, public core plus private harder extension, is one instance of the An old benchmark, reused, no longer measuring what it used to connective theme's broader pattern of the corpus repurposing existing benchmarks until models saturate them.