Training Generative AI GPT-Neo 125M Model

Training GPT-Neo 125M in a Podman Compose Container Using Open Source Tools and Data

If you’re looking to get started with large language models (LLMs) like GPT-Neo 125M, you’re in the right place. In this tutorial, we’ll explore how to train the open source GPT-Neo 125M model from Hugging Face inside a containerized Podman Compose environment.

This setup is perfect for developers who prefer open source tools and want a simple, reproducible training workflow — even on a modest machine.

What Is GPT-Neo 125M?

GPT-Neo 125M is an open source transformer-based language model created by EleutherAI. It’s a great starting point for fine-tuning with small datasets or learning the basics of language model training.

Why Use Podman Compose?

Podman Compose offers a container-based environment similar to Docker Compose — but with a focus on being daemonless and rootless. It’s ideal for developers working on Linux who want an open source alternative to Docker.

With Podman Compose, you can:

• Reproducibly build and run training environments
• Mount volumes for data, model caching, and output
• Avoid polluting the host system

Installing GPT-Neo 125M in Podman Compose

1. Project Structure

gpt-neo-podman-compose/
│—— data/                    # Your custom dataset (.jsonl)
│—— output/                 # Model outputs (saved model)
│—— cache/                  # Model & pip cache
│—— Dockerfile
│—— podman-compose.yml
│—— requirements.txt
│—— train.py

2. Training Script train.py

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments
import torch

# Load dataset
def format_example(example):
    return {"text": f"### Question: {example['prompt']}\n### Answer: {example['response']}"}

dataset = load_dataset("json", data_files="data/custom_dataset.jsonl")
dataset = dataset.map(format_example)
dataset = dataset["train"].train_test_split(test_size=0.1)
dataset = dataset.remove_columns(["prompt", "response"])

# Tokenizer & Model
model_name = "EleutherAI/gpt-neo-125M"  # Smollm 135M is not public, use similar
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

def tokenize(example):
    return tokenizer(example["text"], padding="max_length", truncation=True, max_length=512)

tokenized_dataset = dataset.map(tokenize, batched=True)
tokenized_dataset.set_format(type='torch', columns=['input_ids', 'attention_mask'])

# Training
training_args = TrainingArguments(
    output_dir="./output",
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    num_train_epochs=3,
    logging_steps=10,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    logging_dir="./logs",
    save_total_limit=1,
    fp16=torch.cuda.is_available(),
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset["train"],
    eval_dataset=tokenized_dataset["test"],
    tokenizer=tokenizer,
)

trainer.train()

3. Dockerfile

FROM python:3.10

ENV HF_HOME=/cache/huggingface
ENV TRANSFORMERS_CACHE=/cache/huggingface/transformers
ENV HF_DATASETS_CACHE=/cache/huggingface/datasets
ENV PIP_CACHE_DIR=/cache/pip

WORKDIR /app

# Install system packages
RUN apt update && apt install -y git

# Pre-install dependencies to enable caching
COPY requirements.txt .
RUN pip install --cache-dir=$PIP_CACHE_DIR -r requirements.txt

COPY . .

CMD ["python", "train.py"]

4. The requirements.txt File

transformers
datasets
torch

5. Podman Compose YAML

version: "3"

services:
  smollm-trainer:
    build: .
    volumes:
      - ./data:/app/data
      - ./output:/app/output
      - ./cache:/cache             # Cache for Hugging Face and pip
    command: ["python", "train.py"]

6. Example Open Source Dataset

We’re using a simple JSONL format. Here’s an example from our data/custom_dataset.jsonl file:

[json]
{
“prompt”: “Who is the mayor of Toronto?”,
“response”: “As of 2025, the mayor of Toronto is Olivia Chow.”
}
{
“prompt”: “I need a PHP code snippet to connect to a MySQL database.”,
“response”: “

6. Build and Run with Podman Compose

mkdir -p data output cache
podman-compose build
podman-compose up

This downloads the base GPT-Neo model, loads your dataset, and starts training — all inside a secure, isolated container.

7. Inference Example (Optional)

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("./output")
model = AutoModelForCausalLM.from_pretrained("./output")

prompt = "Who is the mayor of Toronto?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Screenshots & Screencast

Before and After: Results

Output Before Training (default GPT-Neo)

Prompt: Who is the mayor of Toronto?
Response: I’m not sure.

Prompt: I need a PHP code snippet to connect to MySQL.
Response: Incoherent or missing

Output After Fine-Tuning

Prompt: Who is the mayor of Toronto?
Response: As of 2025, the mayor of Toronto is Olivia Chow.

Prompt: I need a PHP code snippet to connect to MySQL.
Response:

<?php
$mysqli = new mysqli("localhost", "user", "password", "database");
if ($mysqli-&gt;connect_error) {
    die("Connection failed: " . $mysqli-&gt;connect_error);
}
echo "Connected successfully";
?>

Additional Resources

If you’re just starting out with Python or want to deepen your understanding, check out my beginner-friendly book and course:

📚 Book:
Learning Python (eBook on Amazon)

🎓 Course:
Learning Python (Ojambo Shop)

Need Help?

I’m available for:

• 💻 One-on-one online Python tutorials: https://ojambo.com/contact
• 🤖 GPT-Neo installation, fine-tuning, or migration: https://ojamboservices.com/contact

Feel free to reach out and let’s bring AI to your next project!

Conclusion

With open source models like GPT-Neo and tools like Podman Compose, training your own AI assistant is more accessible than ever. Whether you’re creating a chatbot, a code assistant, or a personal knowledge base — it all starts with training.

Have questions or want to share your setup? Drop a comment below or get in touch!