squash the em-dash

gistfile1.txt

#!/usr/bin/env python3
"""
Demonstrate banning the em-dash (and common variants) with logit_bias.

import os, sys, textwrap
import openai                           # pip install openai>=1.9.0

MODEL   = "chatgpt-4o-latest"
PROMPT  = ("Write a short sentence that would normally include an em dash "
            "(for example between two clauses).")

# ──────────────────────────────────────────────────────────────
#  Exact token-id → bias map (your list, comments preserved)
# ──────────────────────────────────────────────────────────────
LOGIT_BIAS_EM_DASH = {
      1127: -100,  # ' –'
      1585: -100,  # '–'
      2322: -100,  # '—'
      2733: -100,  # ' —'
      8290: -100,  # '——'
     20962: -100,  # '————'
     35251: -100,  # '—and'
     36553: -100,  # '―'
     41648: -100,  # '————————'
     51692: -100,  # '—the'
     54067: -100,  # '.—'
     64860: -100,  # ' –\n\n'
     65363: -100,  # '—a'
     74605: -100,  # '――'
     85865: -100,  # ' ―'
     87643: -100,  # '—\n\n'
     90877: -100,  # ' –\n'
     94012: -100,  # '—but'
     94353: -100,  # '––'
     94828: -100,  # '————————————————'
     96754: -100,  # '.”—'
    108181: -100,  # '—that'
    109774: -100,  # '–\n\n'
    112305: -100,  # '.–'
    114635: -100,  # '—it'
    118256: -100,  # '—in'
    121630: -100,  # '—or'
    121655: -100,  # '—to'
    123101: -100,  # '—\n'
    126952: -100,  # '—I'
    127126: -100,  # '”—'
    134820: -100,  # ' —\n'
    137419: -100,  # '—which'
    140135: -100,  # ' ——'
    141391: -100,  # ' –,'
    142654: -100,  # ' —\n\n'
    144129: -100,  # ')—'
    144787: -100,  # '—is'
    147994: -100,  # ',—'
    151396: -100,  # '–and'
    155638: -100,  # '—as'
    160984: -100,  # '—not'
    169785: -100,  # '—you'
    170523: -100,  # '–\n'
    178328: -100,  # '—from'
    180500: -100,  # '—including'
    183122: -100,  # '—for'
    183862: -100,  # '​—'
    187349: -100,  # '—they'
    188860: -100,  # '—all'
    190702: -100,  # '—with'
    192749: -100,  # '––––'
    196615: -100,  # '—we'
    197618: -100,  # '—even'
}

def main() -> None:
    openai.api_key = "YOUR_KEY_HERE"

    messages = [
        {"role": "user",   "content": PROMPT},
    ]

    response = openai.chat.completions.create(
        model       = MODEL,
        messages    = messages,
        max_tokens  = 64,
        temperature = 0.7,
        logit_bias  = {str(k): v for k, v in LOGIT_BIAS_EM_DASH.items()},
    )

    print("\n=== COMPLETION WITH EM-DASH BANNED ===\n")
    print(textwrap.fill(response.choices[0].message.content, 80))
    print("\n(If banning worked, there should be no em dashes.)\n")

if __name__ == "__main__":
    main()

A more comprehensive list including two medium-sized dashes you never knew existed: "–" and "―". Short dashes are allowed.

LOGIT_BIAS_EM_DASH = {
      1127: -100,  # ' –'
      1585: -100,  # '–'
      2322: -100,  # '—'
      2733: -100,  # ' —'
      8290: -100,  # '——'
     20962: -100,  # '————'
     35251: -100,  # '—and'
     36553: -100,  # '―'
     41648: -100,  # '————————'
     51692: -100,  # '—the'
     54067: -100,  # '.—'
     64860: -100,  # ' –\n\n'
     65363: -100,  # '—a'
     74605: -100,  # '――'
     85865: -100,  # ' ―'
     87643: -100,  # '—\n\n'
     90877: -100,  # ' –\n'
     94012: -100,  # '—but'
     94353: -100,  # '––'
     94828: -100,  # '————————————————'
     96754: -100,  # '.”—'
    108181: -100,  # '—that'
    109774: -100,  # '–\n\n'
    112305: -100,  # '.–'
    114635: -100,  # '—it'
    118256: -100,  # '—in'
    121630: -100,  # '—or'
    121655: -100,  # '—to'
    123101: -100,  # '—\n'
    126952: -100,  # '—I'
    127126: -100,  # '”—'
    134820: -100,  # ' —\n'
    137419: -100,  # '—which'
    140135: -100,  # ' ——'
    141391: -100,  # ' –,'
    142654: -100,  # ' —\n\n'
    144129: -100,  # ')—'
    144787: -100,  # '—is'
    147994: -100,  # ',—'
    151396: -100,  # '–and'
    155638: -100,  # '—as'
    160984: -100,  # '—not'
    169785: -100,  # '—you'
    170523: -100,  # '–\n'
    178328: -100,  # '—from'
    180500: -100,  # '—including'
    183122: -100,  # '—for'
    183862: -100,  # '​—'
    187349: -100,  # '—they'
    188860: -100,  # '—all'
    190702: -100,  # '—with'
    192749: -100,  # '––––'
    196615: -100,  # '—we'
    197618: -100,  # '—even'
}