PerturbReason/PerturbReason_dataset_code

1# Multi-tier reasoning evaluation with structured triplet output2#3# Modules:4#   data_model      – Shared dataclasses, IO, parsing, external KG loading5#   metric_accuracy – Tier-1: 3-way answer accuracy (up/down/unchanged)6#   metric_symbolic – Tier-2: Deterministic symbolic evaluation (M1-M5)7#   metric_bleurt   – Tier-3: BLEUrT textual similarity8#   metric_llm      – Tier-4: LLM Rescue for unsolved cases9#   pipeline        – Orchestrator that runs all tiers and aggregates10#   runner          – CLI entry-point11 12__version__ = "3.0.0"13