Exploring the “logit_bias” API Parameter: How I Eliminated Em Dashes and Suppressed 106 Tokens — Findings and Code for Your Own “No Dash” Response Test

Mastering Em Dashes in ChatGPT: How to Suppress Them Using logit_bias

Are you tired of seeing unwanted em dashes popping up in your AI-generated responses? If so, you’re not alone. Many users have struggled with plugin, prompt, and memory tweaks that just don’t quite do the trick. However, there’s an effective workaround that involves the powerful logit_bias parameter in OpenAI’s API—an approach that can significantly reduce or even eliminate em dashes from the AI’s output.

The Challenge of Em Dashes

Em dashes—those long, dash-like punctuation marks—add flair but often disrupt formatting or stylistic consistency, especially when precise control over text is required. Users attempting to suppress them have faced disappointing results, as the model stubbornly continues to use em dashes despite custom instructions and context manipulations.

Enter logit_bias: A Targeted Solution

The logit_bias parameter allows developers to influence token probabilities by assigning bias values between -100 and 100. Setting a bias of -100 to specific tokens effectively discourages the model from producing them.

Initially, you might think merely identifying the token id for the em dash (—) is enough. But, as it turns out, the model combines tokens into larger chunks—so simply blocking one token doesn’t prevent variations or similar symbols, like en dashes or hyphens, from appearing.

How I Implemented It

Through experimentation, I discovered that suppressing em dash-related tokens took multiple iterations:

• Step 1: Bias tokens containing the em dash itself.
• Step 2: Accounted for tokens with adjacent characters, like spaces or letters touching the dash.
• Step 3: Addressed similar Unicode symbols, such as en dashes.
• Step 4: Broadened the bias to hyphens used in contexts resembling em dashes, especially when the hyphen is not next to a letter on both sides.

In total, I found that applying a logit_bias of -100 to 106 tokens effectively minimized or eradicated em dash usage in the responses.

Proof in Practice

To illustrate, I used the approach in prompts requesting opinions or solutions on complex topics. Despite model variations with custom instructions and no memories, responses showing the “anti-dash” bias consistently outperformed standard responses:

Sample Prompt:
*”In a paragraph, give me your best hot