Utilizing the “logit_bias” Parameter in the API to Combat Em Dashes: How I Managed to Suppress 106 Tokens and a Guide to Creating Your Own “No Dash” Response Test

Using Logit Bias to Suppress Em Dashes in AI Responses: A Practical Experiment

Dealing with unwanted em dashes in AI-generated text can be a frustrating challenge. Despite attempts with custom instructions and memory tweaks, the presence of em dashes persisted. To address this, I turned to the logit_bias parameter available in the OpenAI API—a powerful way to influence token probability during generation.

The Technical Approach

The core idea was straightforward: identify the token IDs corresponding to the em dash (—) and related characters, then assign a strong negative bias (-100) to effectively ban their usage. However, the process revealed a nuanced complication: tokens for symbols like em dashes can combine with surrounding characters, creating multiple related tokens that also need suppression.

Iterative Token Suppression

It took several rounds of biasing to reach the desired outcome:

• Initially, I targeted tokens immediately representing the em dash—about 3 tokens.
• Then, expanded the scope to include all tokens containing the em dash, roughly 40 tokens.
• When the model started substituting en dashes, the biasing was extended to those as well.
• Finally, with around 106 tokens biased, the model significantly reduced em dash usage, often replacing them with hyphens or avoiding them entirely.

Assessment and Results

I evaluated the impact by prompting various GPT models with a simple “hot take” question, comparing responses with and without the biasing fix. Interestingly, the responses with logit_bias applied maintained overall coherence and quality, especially with more sophisticated models like GPT-4 (the latest versions). The biased models tended to favor more straightforward, dash-free phrasing, aligning with the intended effect.

Key Findings

• Suppressing em dash tokens at the token level doesn’t substantially compromise response quality.
• Even a brute-force approach—applying bias to over 100 tokens—can effectively remove em dashes.
• Fine-tuning models might be a more elegant solution, but this method offers a quick, accessible workaround.

Practical Implementation

I compiled a list of the 106 tokens encountered and created a Python script to apply the negative bias. You can use this script to experiment with your own prompts. Just ensure you set your OPENAI_API_KEY environment variable and make the script executable.

See the full code and token list here:
[Link to GitHub Gist](https://gist.github.com/kernkraft235