Parameters Reference

Complete guide to all configuration parameters available when testing AI models.

---

Overview

Control AI model behavior through these key parameters:

Parameter	Range	Default	Impact
Temperature	0.0 - 2.0	0.7	Creativity vs consistency
Max Tokens	1 - 4096	500	Response length limit
Top P	0.0 - 1.0	1.0	Token selection diversity
Frequency Penalty	-2.0 - 2.0	0.0	Reduce repetition
Presence Penalty	-2.0 - 2.0	0.0	Encourage topic diversity

---

Temperature

Controls randomness/creativity of responses

Range: 0.0 - 2.0

0.0 ━━━━━━━━━━━ 0.7 ━━━━━━━━━━━ 2.0
Deterministic    DEFAULT    Random

When to Use

• 0.0 - 0.3: Facts, data extraction, classification
• 0.4 - 0.9: General use, customer support, content
• 1.0 - 2.0: Creative writing, brainstorming

See: Temperature Guide for detailed guidance

---

Max Tokens

Limits the maximum length of the response

Range: 1 - 4096 (varies by model)

Understanding Tokens:

• 1 token ≈ 4 characters
• 1 token ≈ 0.75 words
• 100 tokens ≈ 75 words

Common Settings

Setting	Tokens	Words	Use Case
Short	100	~75	Quick answers, classifications
Medium	500	~375	Standard responses
Long	1000	~750	Detailed explanations
Very Long	2000	~1500	Articles, reports

Example

{
  "maxTokens": 500,
  "prompt": "Explain quantum computing"
}

Response: Will stop at ~375 words or 500 tokens, whichever comes first.

Best Practices

✅ Set appropriate limits:

• Short answers: 100-300 tokens
• Standard: 300-800 tokens
• Detailed: 800-2000 tokens

✅ Consider cost:

• Fewer tokens = lower cost
• Only pay for tokens you need

❌ Avoid:

• Setting too low (cuts off mid-sentence)
• Setting unnecessarily high (wastes cost)

---

Top P (Nucleus Sampling)

Controls diversity of token selection

Range: 0.0 - 1.0

How It Works:

• Selects from the top P probability mass
• Alternative to temperature
• More predictable than temperature at extremes

Top P = 1.0: Consider all tokens
Top P = 0.9: Consider top 90% probability mass
Top P = 0.5: Consider top 50% (most likely tokens)
Top P = 0.1: Consider top 10% (very focused)

When to Use

Top P	Behavior	Use Case
1.0	All tokens considered	Maximum diversity
0.9	Top 90%	DEFAULT - Balanced
0.5	Top 50%	Focused responses
0.1	Top 10%	Very conservative

Example

{
  "temperature": 0.8,
  "topP": 0.9
}

Result: Creative but still reasonable

Temperature vs Top P

General Rule: Use one or the other, not both at extreme values

// ✅ Good combinations
{temperature: 0.7, topP: 1.0}
{temperature: 1.0, topP: 0.9}

// ❌ Avoid
{temperature: 0.0, topP: 0.1}  // Too restrictive
{temperature: 2.0, topP: 1.0}  // Too random

---

Frequency Penalty

Reduces repetition of tokens based on frequency

Range: -2.0 - 2.0

How It Works:

• Positive: Decreases likelihood of repeated tokens
• Negative: Increases likelihood of repeated tokens
• 0: No penalty (default)

-2.0 ━━━━━━━━━ 0.0 ━━━━━━━━━ 2.0
More repetition  DEFAULT  Less repetition

When to Use

Value	Effect	Use Case
0.0	No penalty	DEFAULT
0.5 - 1.0	Gentle reduction	Avoid minor repetition
1.0 - 2.0	Strong reduction	Prevent repetition
-0.5 - -1.0	Encourage patterns	Technical writing, lists

Example

{
  "frequencyPenalty": 0.5,
  "prompt": "List the benefits of AI"
}

Without Penalty:

AI offers benefits like efficiency. The efficiency of AI...
AI provides efficiency... The efficiency benefits...

With Penalty (0.5):

AI offers benefits like efficiency, accuracy, and scalability.
These advantages include cost reduction and improved decision-making...

Best Practices

✅ Use 0.3 - 0.8 to reduce repetition
✅ Increase if you see repeated phrases
❌ Don't use extreme values unless necessary

---

Presence Penalty

Encourages topic diversity based on presence

Range: -2.0 - 2.0

How It Works:

• Positive: Penalizes tokens that have appeared
• Negative: Rewards tokens that have appeared
• 0: No penalty (default)

Difference from Frequency Penalty:

• Frequency: "How often?" (counts occurrences)
• Presence: "Has it appeared?" (binary)

-2.0 ━━━━━━━━━ 0.0 ━━━━━━━━━ 2.0
Narrow focus    DEFAULT   Diverse topics

When to Use

Value	Effect	Use Case
0.0	No penalty	DEFAULT
0.3 - 0.8	Gentle diversity	Explore related topics
0.8 - 2.0	Strong diversity	Brainstorming, ideation
-0.5 - -1.0	Focus on theme	Stay on topic

Example

{
  "presencePenalty": 0.6,
  "prompt": "Discuss AI applications"
}

Without Penalty:

AI in healthcare... healthcare applications... healthcare benefits...
more healthcare examples... healthcare future...

With Penalty (0.6):

AI in healthcare, education, finance, and manufacturing.
Each sector benefits differently... Transportation uses AI for...

Best Practices

✅ Use 0.3 - 0.8 for topic diversity
✅ Combine with frequency penalty for best results
❌ Extreme values can make responses incoherent

---

Stop Sequences

Defines tokens where the model should stop generating

Format: Array of strings

Use Cases:

• End at specific markers
• Control output structure
• Prevent unwanted continuation

Example

{
  "stopSequences": ["\n\n", "###", "END"],
  "prompt": "List 3 items:\n1."
}

Stops at: Double newline, ###, or END

Common Stop Sequences

// Stop at double newline
stopSequences: ["\n\n"]

// Stop at section markers
stopSequences: ["###", "---", "END"]

// Stop at specific phrases
stopSequences: ["The End", "Conclusion:", "Summary:"]

---

Parameter Combinations

For Different Use Cases

Factual Q&A

{
  "temperature": 0.2,
  "maxTokens": 300,
  "topP": 1.0,
  "frequencyPenalty": 0.0,
  "presencePenalty": 0.0
}

Customer Support

{
  "temperature": 0.7,
  "maxTokens": 500,
  "topP": 0.9,
  "frequencyPenalty": 0.3,
  "presencePenalty": 0.2
}

Creative Writing

{
  "temperature": 1.0,
  "maxTokens": 1000,
  "topP": 0.95,
  "frequencyPenalty": 0.5,
  "presencePenalty": 0.6
}

Code Generation

{
  "temperature": 0.3,
  "maxTokens": 800,
  "topP": 1.0,
  "frequencyPenalty": 0.0,
  "presencePenalty": 0.0
}

---

Testing Strategy

Step 1: Start with Defaults

{
  "temperature": 0.7,
  "maxTokens": 500,
  "topP": 1.0,
  "frequencyPenalty": 0.0,
  "presencePenalty": 0.0
}

Step 2: Adjust One at a Time

If responses are:

• Too random → Decrease temperature
• Too repetitive → Increase frequency penalty
• Too focused → Increase presence penalty
• Too long → Decrease max tokens

Step 3: Fine-Tune

Test variations and measure:

• Quality
• Consistency
• Cost
• User satisfaction

---

Quick Reference

Parameter Cheat Sheet

TEMPERATURE        FREQUENCY PENALTY     PRESENCE PENALTY
0.0-0.3: Facts     0.0: Default          0.0: Default
0.4-0.9: General   0.3-0.8: Reduce reps  0.3-0.8: Topic diversity
1.0-2.0: Creative  >1.0: Strong avoid    >1.0: Explore widely

MAX TOKENS         TOP P                 STOP SEQUENCES
100-300: Short     0.1-0.5: Focused      Custom markers
300-800: Standard  0.9-1.0: Diverse      End indicators
800-2000: Long     1.0: All considered   Section breaks

---

Related Documentation

• Temperature Guide - Deep dive into temperature
• Model Comparison - Choose the right model
• Best Practices - Testing tips
• Sample Prompts - Ready-to-use examples

---

Master these parameters to fine-tune AI responses for your specific needs.