postprocess.py

"""Serving-time post-processing for the frozen persona (35B main_v5 + prompt v0).
Rule from the owner: on every second reply, the crying emoji, which the model uses as sentence punctuation, becomes a period.
Usage: from postprocess import Postprocessor; pp = Postprocessor(); text = pp(text)"""
import re
EMO = '😭'
def emoji_to_period(text: str) -> str:
    # 😭 followed by end/newline/space+capital/space+lowercase: it closes a sentence -> "."
    # 😭 directly before existing punctuation: just drop it. Preceding space is absorbed.
    text = re.sub(r'\s*😭+(?=\s*[.!?,])', '', text)                 # "...money 😭." -> "...money."
    text = re.sub(r'\s*😭+(?=\s*$|\s*\n)', '.', text)                # end of text or line -> "."
    text = re.sub(r'\s*😭+\s+(?=\S)', '. ', text)                     # mid-text -> ". next"
    text = re.sub(r'\.\s*\.', '.', text)                              # collapse doubles
    # capitalize the word after a period we inserted, when the model wrote it lowercase mid-line
    return re.sub(r'(\. )([a-z])', lambda m: m.group(1) + m.group(2).upper(), text)
class Postprocessor:
    def __init__(self): self.n = 0
    def __call__(self, text: str) -> str:
        self.n += 1
        return emoji_to_period(text) if self.n % 2 == 0 else text
if __name__ == '__main__':
    import json, sys
    rows = [json.loads(l) for l in open(sys.argv[1])]
    ctx = [r['reply'] for r in rows if r['kind'] == 'ctx' and r.get('expect') == 'slang_on' and EMO in r['reply']]
    for t in ctx[:4]: print('BEFORE:', t.strip()[:230], '\nAFTER: ', emoji_to_period(t).strip()[:230], '\n')