Experimenting with the API’s “logit_bias” Parameter to Eliminate Em Dashes: How I Suppressed 106 Tokens and Created a “No Dash” Response Comparison—My Results and Code

How to Suppress Em Dashes in GPT Responses Using the OpenAI API’s Logit Bias Parameter

If you’ve ever struggled with ChatGPT’s persistent use of em dashes (—) in its responses, you’re not alone. Despite attempts across various prompts and instructions, the model often defaults to inserting em dashes, which can frustrate those seeking cleaner or more controlled output. Fortunately, there’s an effective workaround using the logit_bias parameter in the OpenAI API.

Understanding the Challenge

Em dashes, along with en dashes and hyphens, are naturally represented as tokens in GPT’s tokenization system. When requesting responses, the model may repeatedly choose to insert these symbols due to their frequency and context within its training data. Custom instructions and memory adjustments often have limited success.

A Targeted Solution: Using logit_bias

The logit_bias parameter allows fine-tuned control over token selection by assigning biases to specific token IDs. Setting a bias value of -100 effectively suppresses that token from the model’s output. The challenge lies in identifying all relevant token IDs associated with em and en dashes, including their potential combinations and variations that the model might produce.

The Process of Suppression

Here’s a step-by-step overview of how to disable em dash output:

1. Identify Tokens Related to Dashes
2. Find token IDs corresponding to the em dash (—).

3. Recognize that the model may generate variations such as spaced em dashes (—), or even hyphen-like characters used in place of dashes.

4. Expand Suppression Beyond Single Tokens

5. It’s insufficient to block only the direct token for —. The model may create composite tokens that include the dash with other characters.

6. Through iterative testing, increase the list of tokens (e.g., include tokens with adjacent spaces, hyphens, en dashes, and hyphen combinations).

7. Apply Biases Incrementally

8. Start with a smaller set (e.g., 3 tokens) to test effects.
9. Gradually expand to include all tokens that encompass or resemble dashes.

10. In the original experiment, it took applying biases to as many as 106 tokens to significantly reduce or eliminate dash usage.

11. Use the ND List for Suppression

12. With the compiled list of token IDs, craft a logit_bias dictionary that assigns each