Output Control: The Art of Brevity

Users don't always want a 5-paragraph answer for a "Yes" or "No" question. Local models, by default, love to explain themselves. This wastes your CPU/GPU cycles and frustrates users.

Here is how to take control of the model's verbosity.

1. The Verbosity Scale

• Verbose: "Sure, I can help with that. The answer to your question is..."
• Moderate: "The answer is X. This happens because of Y."
• Terse: "Answer: X."

---

2. Setting Word and Token Limits

LLMs aren't very good at counting (they process tokens, not words).

• Bad: "Answer in exactly 10 words." (It will likely fail).
• Good: "Answer in 1 sentence." or "Be as brief as possible."

---

3. Using "Post-Processing" instructions

Tell the model what to remove from its output.

• "Provide the answer. Do not provide an introduction or an outro."
• "Provide the SQL code ONLY. No text around it."

---

4. The num_predict Parameter

This is a technical guardrail you can set in a Modelfile (Module 5). PARAMETER num_predict 50

This tells Ollama: "No matter what the model wants, cut it off after 50 tokens." This is a "Hard Limit." It’s perfect for generating titles, category names, or sentiment tags where you know the answer should be very short.

---

5. Controlling Tones

You can control output by defining the "Level of Complexity."

• "Explain it like I am 5 years old." (Short, simple words).
• "Provide a technical summary for a PhD." (Long, complex data-rich sentences).

---

Key Takeaways

• Brevity saves hardware resources and improves user experience.
• Descriptions of length ("1 sentence") are better than number counts ("10 words").
• Use negative constraints to remove "introductions" and "fillers."
• Use the num_predict parameter in your Modelfile for guaranteed short outputs.