Parameter Guide - Quick Defaults¶
Don't overthink parameters! Use these proven defaults and adjust only if needed.
🎯 The 3 Parameters You Actually Need¶
1. Temperature (Creativity)¶
Controls randomness in output. This is the most important parameter.
| Value | Behavior | Use Case |
|---|---|---|
| 0.1-0.3 | Predictable, consistent, factual | Code generation, technical support, data extraction |
| 0.6-0.8 | Balanced, natural conversation | General chatbots, Q&A, tutoring |
| 1.0-1.5 | Creative, varied, surprising | Creative writing, brainstorming, storytelling |
Default recommendation: 0.7
2. Context Window (num_ctx)¶
How much conversation history the model remembers (in tokens).
| Value | Behavior | Use Case |
|---|---|---|
| 2048 | 2-3 exchanges | Quick Q&A, simple tasks |
| 4096 | 5-10 exchanges | Normal conversations (recommended) |
| 8192 | Long discussions | Complex analysis, code review |
Default recommendation: 4096
3. Top-P (Nucleus Sampling)¶
Controls output diversity. Keep this simple.
| Value | Behavior |
|---|---|
| 0.9 | Good default for almost everything |
| 1.0 | Maximum diversity (use with low temperature) |
Default recommendation: 0.9
📋 Copy-Paste Templates¶
General Chatbot (Most Common)¶
PARAMETER temperature 0.7
PARAMETER num_ctx 4096
PARAMETER top_p 0.9
Code Assistant¶
PARAMETER temperature 0.3
PARAMETER num_ctx 8192
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.1
Customer Support Bot¶
PARAMETER temperature 0.3
PARAMETER num_ctx 4096
PARAMETER top_p 0.9
Creative Writer¶
PARAMETER temperature 1.2
PARAMETER num_ctx 8192
PARAMETER top_p 0.95
Translator¶
PARAMETER temperature 0.2
PARAMETER num_ctx 4096
PARAMETER top_p 0.9
Data Extraction / Structured Output¶
PARAMETER temperature 0.1
PARAMETER num_ctx 2048
PARAMETER top_p 0.9
🔧 Optional Parameters (Advanced)¶
These are fine-tuning knobs. Skip unless you have specific needs.
top_k (Token Selection Pool)¶
- Default: 40
- Higher = more diverse output
- Lower = more focused output
- Range: 1-100
PARAMETER top_k 40
repeat_penalty (Repetition Control)¶
- Default: 1.1
- Higher = less repetition (can make output weird)
- Lower = more repetition (natural but redundant)
- Range: 0.5-2.0
PARAMETER repeat_penalty 1.1
stop (Stop Sequences)¶
Stop generation when these strings appear.
PARAMETER stop "User:"
PARAMETER stop "###"
num_predict (Max Response Length)¶
Maximum tokens in response.
PARAMETER num_predict 512
seed (Reproducibility)¶
Set a seed for consistent outputs (useful for testing).
PARAMETER seed 42
🎨 Temperature Visual Guide¶
Temperature Scale:
0.0 ───────────────────────── 2.0
│ │ │
Robotic Natural Chaotic
Examples at different temperatures:
Question: "What is 2+2?"
| Temp | Response |
|---|---|
| 0.1 | "2+2 equals 4." |
| 0.7 | "The answer to 2+2 is 4. This is basic arithmetic." |
| 1.5 | "Well, if we're talking standard mathematics, 2+2 gives us 4, though in some abstract contexts..." |
See the difference? Lower = direct, higher = elaborate.
🧪 How to Experiment¶
Step 1: Start with defaults¶
PARAMETER temperature 0.7
PARAMETER num_ctx 4096
PARAMETER top_p 0.9
Step 2: Test your model¶
make chat
# Select your model and test with real prompts
Step 3: Adjust ONE parameter at a time¶
If output is too boring/repetitive: - Increase temperature by 0.2
If output is too random/incoherent: - Decrease temperature by 0.2
If model "forgets" earlier conversation: - Increase num_ctx to 8192
If responses are too long:
- Add PARAMETER num_predict 256
Step 4: Re-create the model¶
# Edit your Modelfile with new parameters
bash scripts/create-custom-model.sh my-model ./models/custom/my-model/Modelfile
💡 Common Issues and Fixes¶
"Model gives the same response every time"¶
Problem: Temperature too low Fix: Increase to 0.7 or higher
"Model output is nonsense/incoherent"¶
Problem: Temperature too high Fix: Decrease to 0.7 or lower
"Model forgets what I said earlier"¶
Problem: Context window too small Fix: Increase num_ctx to 8192
"Model repeats phrases constantly"¶
Problem: Repetition penalty too low
Fix: Set repeat_penalty 1.2
"Responses are cut off mid-sentence"¶
Problem: Max response length too short
Fix: Add PARAMETER num_predict 1024
📊 Parameter Comparison Table¶
| Use Case | Temp | num_ctx | top_p | repeat_penalty |
|---|---|---|---|---|
| Chatbot | 0.7 | 4096 | 0.9 | 1.1 |
| Code | 0.3 | 8192 | 0.9 | 1.1 |
| Support | 0.3 | 4096 | 0.9 | 1.0 |
| Creative | 1.2 | 8192 | 0.95 | 1.0 |
| Translator | 0.2 | 4096 | 0.9 | 1.0 |
| Data Extract | 0.1 | 2048 | 0.9 | 1.0 |
🚀 Quick Start Recommendations¶
Just want to get started? Use this:
FROM llama3.2:1b
PARAMETER temperature 0.7
PARAMETER num_ctx 4096
PARAMETER top_p 0.9
SYSTEM """
Your instructions here...
"""
This works for 80% of use cases!
Adjust temperature after testing: - Too predictable? → Raise to 0.9 - Too random? → Lower to 0.5
📖 Full Parameter Reference¶
For complete parameter documentation, see Modelfile Reference.
🎓 Learning Path¶
- Week 1: Use defaults (temp 0.7, ctx 4096, top_p 0.9)
- Week 2: Experiment with temperature only
- Week 3: Try different context windows
- Week 4: Explore advanced parameters
Don't try to optimize everything at once!
🔗 Next Steps¶
- Apply these settings: Example Modelfiles
- Understand the concepts: Essential Concepts
- Complete reference: Modelfile Reference
Remember: Good prompts matter more than perfect parameters! Spend your time improving your SYSTEM prompt, not endlessly tweaking numbers.