FMG-Bench: 120 Scenarios for Evaluating LLM Behavior on Theological Questions

FMG-Bench v1 is a corpus of 120 base scenarios with 37 perturbation variants for evaluating large language models. The benchmark focuses on theological triage and pastoral-guidance-adjacent contexts. It was created by FideAI and last updated on HuggingFace in May 2026.

Use Cases

• Benchmarking LLM responses to theological questions based on the described 120 scenarios.
• Studying model behavior on moral guidance and user preference based on perturbation variants.
• Evaluating AI grounding and escalation boundaries in pastoral-adjacent contexts as described.

Strengths

• Contains 120 distinct base scenarios for evaluation.
• Includes 37 perturbation variants for each scenario, allowing for robustness testing.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and overall dataset size are unknown, which may limit suitability assessment.

Provenance

Source

FideAI via HuggingFace.

Collection Method

Likely constructed for benchmark purposes; specific gathering method is not detailed.

Freshness

Last updated 2026-05-27 05:22:11; freshness should be verified.