Heuristic:Arize ai Phoenix Warning Deprecated HallucinationEvaluator
| Knowledge Sources | |
|---|---|
| Domains | Evaluation, Deprecation |
| Last Updated | 2026-02-14 06:00 GMT |
Overview
Deprecation warning: HallucinationEvaluator is deprecated and will be removed in a future version. Use FaithfulnessEvaluator instead.
Description
The HallucinationEvaluator class in phoenix.evals.metrics.hallucination has been fully deprecated. It issues a runtime DeprecationWarning when instantiated. The replacement is FaithfulnessEvaluator, which uses updated terminology:
- Labels changed from factual/hallucinated to faithful/unfaithful
- Score semantics reversed: FaithfulnessEvaluator maximizes score (1.0 = faithful) instead of minimizing it (0.0 = factual)
The module-level docstring, class-level RST docstring, and runtime warning all confirm this deprecation.
Usage
This warning applies when you encounter or consider using HallucinationEvaluator. Migrate all usage to FaithfulnessEvaluator immediately.
The Insight (Rule of Thumb)
- Action: Replace all HallucinationEvaluator usage with FaithfulnessEvaluator
- Value: Score interpretation changes: 1.0 now means "faithful" (good), not "factual" (confusing)
- Trade-off: Label names change, so downstream consumers of evaluation results must update label-matching logic
Reasoning
The Arize Phoenix team deprecated HallucinationEvaluator to adopt clearer terminology. The term "hallucination" conflates multiple failure modes, while "faithfulness" directly describes whether model output is grounded in provided context. The score direction change (maximize = good) aligns with standard ML conventions.