Spaces:
Running
Running
| import re | |
| def evaluate_fluency(llm_response: str) -> dict: | |
| issues = [] | |
| if len(llm_response.split()) < 5: | |
| issues.append("Response too short") | |
| if len(llm_response.split()) > 500: | |
| issues.append("Response too long") | |
| # Find the first alphabetic character, skipping list markers like | |
| # "1.", "-", "*", "•", whitespace, etc. A numbered/bulleted answer | |
| # starting with "1. The system..." is still properly capitalized — | |
| # it just isn't index [0]. | |
| first_letter_match = re.search(r'[a-zA-Z]', llm_response) | |
| if first_letter_match and not first_letter_match.group().isupper(): | |
| issues.append("Does not start with capital letter") | |
| # Strip trailing whitespace before checking the last character, and | |
| # allow a closing bullet/list item that ends in a letter followed by | |
| # nothing (common in list-formatted answers) without penalizing it | |
| # as harshly -- but still flag genuinely unpunctuated prose. | |
| stripped = llm_response.rstrip() | |
| if stripped and stripped[-1] not in [".", "?", "!", ":"]: | |
| issues.append("Does not end with proper punctuation") | |
| words = llm_response.lower().split() | |
| for i in range(len(words) - 1): | |
| if words[i] == words[i + 1]: | |
| issues.append("contains repeated consecutive words") | |
| break | |
| if len(issues) == 0: | |
| verdict = "Fluent" | |
| elif len(issues) <= 2: | |
| verdict = "Partially Fluent" | |
| else: | |
| verdict = "Not Fluent" | |
| return { | |
| "issues": issues, | |
| "verdict": verdict | |
| } |