4.1 Deep Dive: Master LLM Hyperparameters with 6 Production-Ready Python Profiles

When developing applications with Large Language Models (LLMs), passing raw prompt text is only half the battle. To build predictable, enterprise-grade software, you must know how to manipulate the underlying mechanics of the API. By tuning hyperparameters like temperature, penalties, and structure constraints, you transform an unpredictable chat model into a reliable software component.

Here is your definitive field guide to the core LLM parameters, followed by 6 distinct code configurations tailored for real-world application workloads, including the exact prompts and responses for each.

The LLM Control Panel: Understanding the Dials

Behind the scenes, an LLM calculates a probability score for every single word (token) in its dictionary. These settings act like dials on a mixing board, dynamically forcing the API to alter its selection behavior:

Temperature: Controls randomness and creativity. Setting it to 0.0 forces the model to choose only the most mathematically probable word every single time, making outputs highly deterministic. Raising it toward 1.0 or higher opens the selection pool to less probable words, boosting creative variation.
Top_p (Nucleus Sampling): An alternative way to control randomness. A top_p of 0.3 means the model permanently discards the bottom 70% least likely words, choosing only from the top 30% safest options. (Best practice: Tune either Temperature or Top_p, never both at the same time).
Max Tokens: Hardcodes a ceiling on the length of the model's response to keep costs predictable and prevent runaway generations.
Stop Sequences: A custom string or list of strings that acts as an instant cut-off switch. The moment the AI generates this specific text pattern, it stops writing immediately.
Frequency Penalty: Punishes tokens based on how many times they have already appeared in the text. Bumping this up forces the model to use a broader vocabulary and stops it from repeating its own catchphrases.
Presence Penalty: Punishes tokens simply for having appeared at least once. This forces the model to constantly jump to completely fresh topics and ideas.

1. The Parameters Explicitly in your PDFs

Your slides explicitly cover the math and theory behind these standard features:

temperature (Lesson 1, Pages 32–37): Discussed as the core control for predictability vs. creativity.
top_p (Lesson 1, Pages 38–40): Defined as nucleus sampling to control vocabulary selection pools.
frequency_penalty & presence_penalty (Lesson 1, Pages 43–47): Taught as the foundational methods to prevent repetition and encourage subject diversity.

2. The Code Parameters NOT in your PDFs (New for your Blog!)

To build functional Python code, we introduced newer, production-grade OpenAI SDK parameters that your lesson slides do not include:

max_completion_tokens: Your slides still refer to the older, legacy parameter max_tokens. In modern production environments, OpenAI uses max_completion_tokens to ensure newer reasoning models account for both internal logic steps and the final visible output.
response_format={"type": "json_object"}: Enforcing a strict JSON format is an advanced API engineering setting that is not covered in your prompting lessons.
n: The exact parameter that instructs the server to fork its thinking and return multiple choices simultaneously is missing from the slides. Your lessons teach you the theory of running multiple paths for Self-Consistency but don't show you the n parameter used to execute it efficiently in code [Part_2].
stop: The programmatic use of cut-off character lists (like stop=["."]) is completely absent from the file text.

6 Production Code Profiles (Using the OpenAI Python SDK)

To implement these in your codebase, initialize your client connection first:

Python

import openai

client = openai.OpenAI()

Profile 1: The Soulless Corporate Robot (Strict Determinism)

Best Used For: Mathematics, code translation, data migrations, and any logic task where zero variation or random hallucination can be tolerated.

Python

def get_robotic_completion(prompt, model="gpt-5-mini"):

response = client.chat.completions.create(

model=model,

messages=[{"role": "user", "content": prompt}],

temperature=0.0, # Eliminates randomness entirely

max_completion_tokens=1000

)

return response.choices.message.content

# --- EXAMPLE WORKLOAD ---

# Prompt: "Translate this sentence into corporate speak: 'The chicken crossed the road because it wanted to get away from the farmer.'"

# Expected Response: "The biological unit classified as 'Chicken' initiated a strategic cross-functional corridor transit. This deployment was executed to mitigate proximity risks associated with the agricultural manager and optimize long-term asset isolation."

Profile 2: The Caffeine-Fueled Brainstormer (High Creativity & Topic Shifting)

Best Used For: Marketing hooks, podcast outlines, and creative copy. It pushes the boundaries of imagination while enforcing penalties so the text stays fresh and dynamic.

Python

def get_chaotic_completion(prompt, model="gpt-5-mini"):

response = client.chat.completions.create(

model=model,

messages=[{"role": "user", "content": prompt}],

temperature=0.9, # Elevates vocabulary variety

frequency_penalty=0.6, # Prevents word repetition

presence_penalty=0.9, # Forces rapid transitions to new concepts

max_completion_tokens=2000

)

return response.choices.message.content

# --- EXAMPLE WORKLOAD ---

# Prompt: "Give me a wild movie pitch about a vintage 1970 Dodge Challenger."

# Expected Response: "Picture this: A classic midnight-blue Challenger R/T is struck by a freak lightning storm. Instead of shifting gears, it rips open rifts in time, but it only runs on high-octane espresso. Suddenly, the driver is selling muscle car insurance to Julius Caesar. Next thing you know, cybernetic dinosaurs show up wearing neon sunglasses. It's a high-speed, reality-bending ride!"

Profile 3: The Short-Fused Bouncer (Hard Structural Boundaries)

Best Used For: Single-word text classification or short-form classification queries. It keeps things highly deterministic and shuts down the generation the millisecond a specific boundary condition is met.

Python

def get_abrupt_completion(prompt, model="gpt-5-mini"):

response = client.chat.completions.create(

model=model,

messages=[{"role": "user", "content": prompt}],

temperature=0.1, # Low variation for safety

max_completion_tokens=15, # Enforces absolute brevity

stop=["."] # Terminates immediately at the first period

)

return response.choices.message.content

# --- EXAMPLE WORKLOAD ---

# Prompt: "Is a loud 'clunking' sound from a front car wheel dangerous? Answer yes or no."

# Expected Response: "Yes"

Profile 4: The Calculated Critic (Nucleus Sampling Mastery)

Best Used For: Long-form summaries or objective critiques. By utilizing top_p instead of temperature, it cleanly eliminates the low-probability junk words from the vocabulary pool, ensuring an authoritative tone.

Python

def get_opinionated_completion(prompt, model="gpt-5-mini"):

response = client.chat.completions.create(

model=model,

messages=[{"role": "user", "content": prompt}],

top_p=0.3, # Discards the bottom 70% of the vocabulary pool

max_completion_tokens=1200

)

return response.choices.message.content

# --- EXAMPLE WORKLOAD ---

# Prompt: "Provide a direct critique of swapping an electric EV motor into a classic 1969 Dodge Charger."

# Expected Response: "This conversion is a fundamental violation of automotive heritage. A classic muscle car relies entirely on the mechanical character and acoustic profile of a V8 internal combustion engine. Removing the powertrain to install a silent battery pack strips away the historical authenticity of the chassis. While it achieves modern efficiency benchmarks, it compromises the core identity of the vehicle."

Profile 5: The Clean Data Pipe (Enforced JSON Mode)

Best Used For: Backend application pipelines where the output must be ingested by a database or another system script without crashing. It forces the neural network to output valid structural syntax.

Python

def get_structured_json_completion(prompt, model="gpt-5-mini"):

response = client.chat.completions.create(

model=model,

messages=[{"role": "user", "content": prompt}],

temperature=0.1,

response_format={"type": "json_object"} # System forces valid JSON structure

)

return response.choices.message.content

# --- EXAMPLE WORKLOAD ---

# Prompt: "Return a JSON object for a 1970 Dodge Challenger R/T. Include the keys 'engine', 'horsepower', and 'quarter_mile_time'."

# Expected Response:

# {

# "engine": "440 Six Pack V8",

# "horsepower": 390,

# "quarter_mile_time": "13.7 seconds"

# }

Profile 6: The Voting Machine (Parallel Candidate Sampling)

Best Used For: Building advanced ensemble checking setups or manual self-consistency scripts. Passing n=3 commands the engine to fork its thinking and return three unique paths at once to maximize execution efficiency.

Python

def get_multi_path_completion(prompt, model="gpt-5-mini"):

response = client.chat.completions.create(

model=model,

messages=[{"role": "user", "content": prompt}],

temperature=0.7, # Provides variation across the paths

n=3 # Instructs the server to return 3 completions at once

)

# Extract all parallel reasoning chains into a single iterable list

paths = [choice.message.content for choice in response.choices]

return paths

# --- EXAMPLE WORKLOAD ---

# Prompt: "A garage holds 5 cars. If the owner buys 2 more enclosed trailers that hold 4 cars each, how many total spots does he have? Show the basic math step."

# Expected Response:

# Returns a list containing 3 unique paths generated simultaneously:

# [

# "Path 1: 2 trailers * 4 cars = 8 spots. 8 spots + 5 garage spots = 13 total spots.",

# "Path 2: Start with 5 spots. Add 4 spots from trailer one, then add 4 spots from trailer two to equal 13.",

# "Path 3: Total capacity equals the original 5 plus the new trailer capacity of 8 (2 x 4), resulting in 13 spots."

# ]

9 AI 101

Saturday, May 23, 2026