First Steps into LLM Training

First Steps into LLM Training

After 20 years in industrial maintenance planning, I did something today that would have seemed like science fiction a decade ago: I trained an artificial intelligence model on my own computer. Not hired someone to do it. Not purchased expensive software. I actually modified the neural pathways of a language model myself.

And honestly? It wasn't nearly as complicated as I expected.

Why a Maintenance Guy is Learning About LLMs

Let me be clear about something upfront: I'm not trying to become a data scientist. I'm not planning to build the next ChatGPT. What I am doing is positioning myself to understand how these tools work well enough to implement them effectively in industrial settings. There's a massive gap right now between what AI can theoretically do and what actually gets deployed successfully in places like refineries, chemical plants, and processing facilities.

I've spent two decades planning shutdowns, managing SAP and Maximo systems, and coordinating maintenance activities across electrical, instrumentation, mechanical, and pipefitting disciplines. I know what works on a plant floor and what doesn't. Now I'm learning the technical side of AI - not to replace my operational knowledge, but to complement it.

Today was my first real hands-on session with Large Language Models (LLMs). I wanted to understand the fundamentals: How do you actually train these things? What does "fine-tuning" really mean? Can someone without a PhD in computer science make meaningful modifications to an AI model?

The answer to that last question is yes. But it requires breaking the learning down into manageable chunks.

The Learning Approach: Small Steps, Real Understanding

My approach to learning LLMs mirrors how I've always approached complex maintenance systems. You don't try to understand an entire refinery's control system in one day. You start with one loop, one instrument, one process. You build understanding incrementally until the bigger picture emerges.

Today's session focused on three things:

1. Understanding how to load and run a pre-trained model

2. Learning what fine-tuning actually means in practice

3. Executing a simple training loop to modify model behavior

We started with GPT-2, an older but well-documented model that's small enough to run on a standard laptop. I loaded it in JupyterLab - a browser-based coding environment that's become standard for data science work. With just a few lines of Python code, I had a working AI generating text.

The first test was simple: give it a prompt and see what it generates. "The future of AI is..." produced a rambling response that was grammatically correct but unfocused. Not impressive by today's standards, but it worked. More importantly, I understood what was happening under the hood.

Fine-Tuning: Teaching an Old Model New Tricks

Here's where it got interesting. Fine-tuning isn't about building a model from scratch - that requires massive computational resources and datasets. Instead, you take an existing model and teach it new patterns by showing it examples.

I created a tiny training dataset - just five examples - teaching GPT-2 to respond like a pirate. Yes, a pirate. It sounds silly, but it perfectly demonstrates the concept. Each training example showed the model what I wanted: normal input, followed by a pirate-style response.

The training process took about three minutes on my laptop. The model processed these examples multiple times (three "epochs" in AI terminology), adjusting its internal parameters slightly each time to match the pattern I was showing it.

Did it work perfectly? No. With only five examples and three training passes, the effect was subtle. But that's exactly the point - I learned what it takes to see meaningful results. Real fine-tuning projects use hundreds or thousands of examples and more sophisticated techniques. But the fundamental process is the same: show the model what you want, let it learn the pattern, test the results.

This incremental learning approach is crucial. I didn't walk away thinking I'm now an AI expert. I walked away understanding one specific piece of the puzzle. Tomorrow's piece will build on today's. That's how expertise develops.

Three Key Lessons for Maintenance Applications

Lesson 1: AI Models Are Tools That Need Training Data - Just Like Technicians Need Procedures

The most important thing I learned today is that AI models are fundamentally pattern recognition systems. They learn from examples. This has direct implications for maintenance applications.

Maintenance Application Example: Imagine training an LLM on your facility's work order history. Every time a pump fails, there's a work order with symptoms, causes, and corrective actions. Feed thousands of these into a model, and it can start recognizing patterns: "High vibration readings plus elevated bearing temperature plus increased power draw usually means bearing failure within 48 hours."

But here's the critical part: the model is only as good as your data. If your work orders are incomplete, inconsistent, or poorly documented, the AI will learn garbage patterns. This isn't magic - it's sophisticated pattern matching. The lesson: before you can effectively deploy AI in maintenance, you need clean, consistent data. That's a process improvement challenge, not a technology challenge.

Lesson 2: Fine-Tuning Allows Customization Without Starting From Scratch

Today I learned you don't need to build a model from the ground up. You take something that already understands language (or images, or time-series data) and teach it your specific domain.

Maintenance Application Example: Take a general-purpose LLM and fine-tune it on your facility's maintenance procedures, safety protocols, and equipment manuals. Suddenly you have an AI assistant that can answer questions like:

• "What's the lockout/tagout procedure for the main cooling water pump?"

• "What torque specs do we use for flange bolts on 6-inch 300# piping?"

• "Show me the troubleshooting steps for high discharge pressure on compressor C-101."

The base model understands language structure and general concepts. Fine-tuning teaches it your specific equipment, procedures, and terminology. This is vastly easier than building a domain-specific AI from scratch, and it's becoming increasingly accessible to non-specialists.

Lesson 3: The Implementation Gap Is Real - And That's Where the Opportunity Is

The technical barrier to working with AI is lower than I expected. Setting up the environment took some guidance, but once configured, running and training models is surprisingly straightforward. The code itself is often just a few dozen lines.

Maintenance Application Example: The real challenge isn't the AI technology - it's the implementation. How do you integrate an AI model with your existing Maximo or SAP system? How do you handle data security in a plant environment? How do you train maintenance planners to use AI-augmented tools effectively? How do you overcome the cultural resistance to "letting the computer make decisions"?

These are organizational and process challenges, not technical ones. And this is where someone with deep operational experience has an advantage. I can't out-code a data scientist, but I can bridge the gap between what AI can theoretically do and what actually works on a plant floor. I understand the constraints, the workflows, the politics, and the practical realities that determine whether a technology gets adopted or abandoned.

What's Next

Today was day one of a much longer journey. I'm not trying to master AI overnight - I'm building understanding piece by piece. The roadmap I've laid out spans 14 months of incremental learning, hands-on projects, and practical application.

The path forward includes:

• Building predictive maintenance models using real equipment failure data

• Creating AI tools that analyze data quality in maintenance management systems

• Developing chatbots that can answer questions from maintenance procedures and technical manuals

• Understanding how to integrate AI capabilities with existing enterprise asset management platforms

Each step builds on the previous one. Each project teaches specific, applicable skills. By this time next year, I won't be an AI researcher, but I will be someone who can effectively guide AI implementation in industrial maintenance settings.

For fellow maintenance professionals looking at AI and wondering where to start: start small. Pick one concept. Build one simple project. Understand one piece of the puzzle. The technology is more accessible than you think, and the industry desperately needs people who can bridge the gap between operational reality and technological possibility.

The future of maintenance isn't about AI replacing human expertise - it's about augmenting it. Understanding how these tools actually work is the first step toward using them effectively.

About the Author: After 20 years in maintenance planning across oil, gas, and chemical industries, I'm spending 2026 learning how to implement AI in industrial settings. Follow along as I document the journey from experienced maintenance professional to AI implementation specialist. Currently enrolled in Python and Data Science programs, targeting consulting work in AI-augmented maintenance by 2027.ntent