LLMBasedFaithfulness fails when response cannot be split into score and reasoning

[kelvinchanwh]

In continuous_eval/metrics/generation/text/llm_based.py line 72-77, the response is split into a score and reasoning. If the split fails, the score in line 76 should check if "yes" is in response.lower() rather than score_txt.lower() since score_txt has not been defined yet.

AS-IS:

try:
    score_txt, reasoning = response.split("\n", 1)
    score = float("yes" in score_txt.lower())
except ValueError:
    score = float("yes" in score_txt.lower())
    reasoning = response

TO-BE:

  try:
      score_txt, reasoning = response.split("\n", 1)
      score = float("yes" in score_txt.lower())
  except ValueError:
      score = float("yes" in response.lower())
      reasoning = response