パラメーターによるモデルの挙動制御

The Core Parameters That Matter

Temperature: Controlling Randomness

from openai import OpenAI
client = OpenAI()

for temp in [0.0, 0.7, 1.5]:
    response = client.chat.completions.create(
        model='gpt-4o-mini',
        messages=[{'role': 'user', 'content': 'Name a color.'}],
        temperature=temp,
        max_tokens=5
    )
    print(f'Temp {temp}: {response.choices[0].message.content}')
# Temp 0.0: Red       (always most common)
# Temp 0.7: Blue      (varied but sensible)
# Temp 1.5: Vermillion (surprising choices)

max_tokens: Controlling Response Length

# Different max_tokens for different use cases

# Classification: short answer expected
classification_response = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=[{'role': 'user', 'content': 'Is this positive or negative? "Great product!"'}],
    max_tokens=5  # Only need 1-2 words
)

# Detailed analysis: longer output needed
analysis_response = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=[{'role': 'user', 'content': 'Analyze the pros and cons of microservices.'}],
    max_tokens=800  # Need space for detailed explanation
)

top_p: Nucleus Sampling

# top_p usage example
response_narrow = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=[{'role': 'user', 'content': 'Continue: The sky is...'}],
    top_p=0.1,  # only very likely tokens (conservative, predictable)
    temperature=1.0  # keep temperature at 1 when using top_p
)

response_wide = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=[{'role': 'user', 'content': 'Continue: The sky is...'}],
    top_p=0.95,  # most tokens eligible (creative, varied)
    temperature=1.0
)

Frequency Penalty: Reducing Repetition

# Frequency penalty to reduce repetition in long outputs
response = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=[{
        'role': 'user',
        'content': 'Write 5 tips for better sleep.'
    }],
    max_tokens=300,
    frequency_penalty=0.5  # reduces repeating the same words/phrases
)
print(response.choices[0].message.content)

Presence Penalty: Encouraging Topic Diversity

# Presence penalty for diverse brainstorming output
response = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=[{
        'role': 'user',
        'content': 'List 10 creative ways to use AI in a small business.'
    }],
    max_tokens=400,
    presence_penalty=0.8  # encourages introducing different topics per item
)
print(response.choices[0].message.content)

stop Sequences: Custom Stopping Points

# Use stop sequences to get clean single-line output
response = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=[{
        'role': 'user',
        'content': 'What is the Python keyword for a function definition?\nAnswer:'
    }],
    max_tokens=20,
    stop=['\n', '.']  # stop at newline or period - gets just the keyword
)
print(repr(response.choices[0].message.content))  # 'def'

seed: Reproducible Outputs

# Reproducible output with seed parameter
response1 = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=[{'role': 'user', 'content': 'Pick a random number from 1 to 10.'}],
    temperature=0,
    seed=42
)

response2 = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=[{'role': 'user', 'content': 'Pick a random number from 1 to 10.'}],
    temperature=0,
    seed=42
)

print(response1.choices[0].message.content)  # same
print(response2.choices[0].message.content)  # same
print('Fingerprint:', response1.system_fingerprint)

Choosing Parameters for Common Use Cases

# Parameter presets for different task types
PRESETS = {
    'classify': {'temperature': 0, 'max_tokens': 20},
    'factual_qa': {'temperature': 0.2, 'max_tokens': 400},
    'creative': {'temperature': 0.9, 'max_tokens': 1000, 'frequency_penalty': 0.3},
    'code': {'temperature': 0.1, 'max_tokens': 2000},
    'summary': {'temperature': 0.3, 'max_tokens': 300, 'frequency_penalty': 0.2},
}

def complete(prompt, task_type='factual_qa', **overrides):
    params = {**PRESETS[task_type], **overrides}
    return client.chat.completions.create(
        model='gpt-4o-mini',
        messages=[{'role': 'user', 'content': prompt}],
        **params
    )

Logprobs: Understanding Model Confidence

import math

response = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=[{'role': 'user', 'content': 'The capital of France is?'}],
    max_tokens=3,
    logprobs=True,
    top_logprobs=3  # show top 3 alternative tokens at each position
)

for token_log in response.choices[0].logprobs.content:
    prob = math.exp(token_log.logprob)  # convert log prob to probability
    print(f'Token: {token_log.token!r} | Probability: {prob:.2%}')
    for alt in token_log.top_logprobs:
        print(f'  Alt: {alt.token!r} -> {math.exp(alt.logprob):.2%}')

Testing Parameter Effects Systematically

from itertools import product

# Systematic parameter grid search
temperatures = [0.0, 0.3, 0.7]
max_tokens_options = [100, 300]
test_prompt = 'Summarize the benefits of unit testing in 2 sentences.'

results = []
for temp, max_tok in product(temperatures, max_tokens_options):
    resp = client.chat.completions.create(
        model='gpt-4o-mini',
        messages=[{'role': 'user', 'content': test_prompt}],
        temperature=temp,
        max_tokens=max_tok
    )
    results.append({
        'temperature': temp,
        'max_tokens': max_tok,
        'output': resp.choices[0].message.content,
        'actual_tokens': resp.usage.completion_tokens
    })

Quick Check

Lesson Recap