Leveraging the “logit_bias” Parameter in the API to Combat Em Dashes: My Experience with Reducing 106 Tokens and a Sample Code for “Dash-Free” Responses (Variation 36)

Mastering Em Dashes in GPT API: How to Suppress Them Using Logit Bias

If you’ve ever struggled with GPT models incessantly inserting em dashes (—) into your prompts, you’re not alone. Many users find it challenging to control or eliminate these punctuation marks, especially when precision matters. Recently, I embarked on an experiment to suppress em dashes by leveraging the logit_bias parameter in the OpenAI API—a powerful yet underused feature.

The Challenge with Em Dashes

I wanted GPT to avoid using em dashes altogether, but standard approaches—like custom instructions or trying to influence memory—proved ineffective. The model kept defaulting to em dashes, even after multiple attempts. That’s when I recalled the logit_bias parameter, which allows us to adjust the probability of specific tokens appearing in the output.

Using logit_bias to Suppress Dashes

Initially, I identified the token ID for the em dash (—). However, because tokenization can produce multiple tokens for a single symbol—especially with punctuation and compound characters—it was necessary to apply biasing more broadly. I systematically found all tokens related to the em dash, including variations with surrounding spaces and different dash types like en dashes and hyphens.

Here’s the progression I followed:

• Step 1: Bias all tokens explicitly representing the em dash (—). This required setting their bias to -100.
• Step 2: Expand to include tokens with characters touching the dash—letters or punctuation adjacent to the dash—totaling around 40 tokens biased.
• Step 3: Account for en dashes and other dash variants by increasing the biasing scope.
• Step 4: Finally, bias hyphen tokens that are used as em dashes in certain contexts, especially those not connected to letters, setting their bias to -100 as well.

Remarkably, setting biases on 106 tokens effectively persuaded GPT-4 to almost completely eliminate em dashes from responses.

Observed Results

Efforts to suppress em dashes did not significantly degrade the quality of responses. Here’s a comparison illustrating the impact:

Prompt: In a paragraph, give me your best ‘hot take’

• Normal Response: You get a typical, well-phrased hot take about productivity culture.
• Bias-Adjusted Response: The response shifts, avoiding em dash characters, sometimes replacing them with hyphens or rephr