File judge_eval.py

benchmarking/vllm/judge_eval.py:None–None

"""LLM-as-a-judge evaluation for GSM8K samples.

lm-eval's regex scoring misses answers in \\boxed{X} format. This script
re-evaluates the JSONL samples using a judge LLM to get real accuracy.