Skip to content

AI Parameter Playbook — Master LLM Parameters (Free & Local-First)

Learn exactly what each LLM parameter does through hands-on experiments on your laptop. No APIs, no cloud costs.

🎓 Start Here: 45-Minute Learning Path

Understand parameters by testing them locally:

  1. Temperature Sweep (5 min) ← How does randomness change?
  2. Top-p Sweep (5 min) ← How do filters restrict candidates?
  3. Top-k Sweep (5 min) ← How do ranked cutoffs work?
  4. Combined Filters (10 min) ← What happens when you mix them?
  5. Repetition Penalties (5 min) ← How do penalties work?
  6. Real Use Cases (15 min) ← How to tune for different tasks?

→ Start the Learning Path (no prior knowledge needed)

Or run the simulator right now (5 minutes):

pip install -r requirements-experiments.txt
python3 experiments/local/sampling_simulator.py

Why Parameters Matter

Every LLM has the same problem: at each step, it generates probabilities for every possible next token. Parameters control how it picks the next token.

flowchart LR
  A["Prompt<br/>(your input)"] --> B["Model predicts<br/>next-token odds"]
  B --> C["Parameter transforms<br/>(temperature, filters, penalties)"]
  C --> D["Sample a token<br/>(pick the next word)"]
  D --> E["Repeat until<br/>done or stop"]
  E --> F["Output<br/>(LLM response)"]

The magic: Different parameters lead to wildly different behaviors on the same prompt.

  • Low temperature = deterministic, safe, boring
  • High temperature = creative, risky, interesting
  • Top-p filters = dynamic nucleus based on confidence
  • Penalties = reduce repetition

This playbook teaches you to control this intentionally.

Three Learning Paths

🎯 Path 1: Quick Intuition (30 minutes)

Result: You'll know what each parameter does and how to tune for common tasks.

🧠 Path 2: Deep Dive (90 minutes)

Complete all 6 experiments in sequence. Build intuition about: - How parameters reshape distributions - How filters interact - How penalties discourage repetition - How to design parameters for different use cases

→ Start here: Learning Path Overview

☁️ Path 3: Cloud Validation (optional)

Once you've mastered parameters locally, test on any cloud platform of your choice:

  • Azure OpenAI – Enterprise compliance focus
  • OpenAI – Latest models (GPT-4, o1)
  • Google Vertex AI – Gemini family, GCP integration
  • Amazon Bedrock – Multiple vendors (Claude, Llama, Mistral)

All platforms support the same core parameters. Learn them once, apply everywhere.

→ Choose Your Cloud Platform

What You'll Know After

✅ What each parameter actually does (math + intuition)
✅ How parameters interact and affect each other
✅ How to tune for different tasks (creative vs. safe)
✅ How to diagnose parameter problems
✅ How to apply this to any LLM (Azure, OpenAI, local, etc.)

Local Playground

No credentials needed. These run entirely offline:

Install experiment dependencies with pip install -r requirements-experiments.txt.

Reference (When You're Ready)

Authoritative References

This playbook is based on official specs:

  • Azure OpenAI inference REST API (GA 2024-10-21)
  • OpenAI Chat Completions API reference
  • Open-source sampler implementations (llama.cpp, Hugging Face)

TL;DR

Your first step:

pip install -r requirements-experiments.txt
python3 experiments/local/sampling_simulator.py  # 2 minutes

Then:

→ Read the Learning Path Overview (5 min)

That's it. 7 minutes and you'll understand LLM parameters deeper than most people.