Runtime Parameters: The Control Panel

In Module 4, we learned the theory of why parameters matter. Now, we will look at how to actually apply them in a Modelfile. These PARAMETER commands allow you to control the "statistics" of how the model picks the next word.

1. Temperature (Creativity)

• Range: 0.0 to 2.0 (Default ~0.8)
• Low (0.1): Predictable. Good for code, math, and data extraction.
• High (1.2): Randomized. Good for poetry, creative writing, and ideas.
• Modelfile: PARAMETER temperature 0.2

---

2. Top-K and Top-P (The Filters)

These and "Temperature" work together to narrow down the model's choices.

• Top-K: Tells the model to only look at the top K most likely words. (PARAMETER top_k 40)
• Top-P: Tells the model to only look at a group of words whose total probability equals P. (PARAMETER top_p 0.9)

Tip: Usually, you only need to change Temperature. Ollama’s default Top-K/P settings are excellent for most users.

---

3. Context Window (num_ctx)

As discussed in Module 4, this is the "memory size."

• Small (2048): Saves RAM, very fast.
• Standard (8192): Good for general chat.
• Large (32768+): For analyzing long documents.
• Modelfile: PARAMETER num_ctx 16384

---

4. Penalty Parameters

Sometimes models get "stuck" in a loop, repeating the same word ("the the the...").

• Repeat Penalty: How hard to punish the model for repeating itself. (PARAMETER repeat_penalty 1.1)
• Presence Penalty: Encourages the model to talk about new topics.

---

5. Stop Sequences (stop)

Important for developers! You can tell Ollama to stop generating text as soon as it sees a specific word.

Example: If you want the model to generate a Python function and then stop without explaining it, you might look for the final ``` block. `PARAMETER stop " ``` "`

---

6. Seed (For Predictability)

If you want the model to give the exact same answer every time you run the same prompt, set the seed to a specific number. PARAMETER seed 42

---

Summary Cheat Sheet

Parameter	Default	Use Case
temperature	0.8	Change "Vibe" vs "Fact"
num_ctx	4096 (usually)	Increase working memory
stop	None	Control API response limits
num_gpu	Auto	Force model to CPU (0) or GPU (1)
repeat_last_n	64	How far back to check for repetition

---

Key Takeaways

• Parameters go inside the Modelfile after the SYSTEM prompt.
• Temperature is the most commonly changed setting.
• num_ctx is the most important setting for hardware stability.
• Use Stop sequences to keep your outputs clean for automation.
• The PARAMETER keyword is followed by the name and the value.