Table of Contents

• Quick Start
• Installation
• Data Preparation
  • Option 1: Default FineWeb Dataset
  • Option 2: Custom Data
• Model Requirements
• Training Guide
  • LLMTrainer Parameters
  • Training Parameters
• Contributing

llm_trainer in 5 Lines of Code

from llm_trainer import create_dataset, LLMTrainer

create_dataset(save_dir="data")   # Generate the default FineWeb dataset
model = ...                       # Define or load your model (GPT, xLSTM, Mamba...)
trainer = LLMTrainer(model)       # Initialize trainer with default settings
trainer.train(data_dir="data")    # Start training on the dataset

🔴 YouTube Video: Train LLMs in code, spelled out

[!NOTE] Explore usage examples

Installation

$ pip install llm-trainer

How to Prepare Data

Option 1: Use the Default FineWeb Dataset

from llm_trainer import create_dataset

create_dataset(save_dir="data",         # Where to save created dataset
               chunks_limit=1_500,      # Maximum number of files (chunks) with tokens to create
               chunk_size=int(1e6))     # Number of tokens per chunk

Option 2: Use your own data

1. Your dataset should be structured as a JSON array, where each entry contains a "text" field. You can store your data in one or multiple JSON files.

Example JSON file:

[
   {"text": "Learn about LLMs: https://www.youtube.com/@_NickTech"},
   {"text": "Open-source python library to train LLMs: https://github.com/Skripkon/llm_trainer."},
   {"text": "My name is Nikolay Skripko. Hello from Russia (2025)."}
]

2. Run the following code to convert your JSON files into a tokenized dataset:

from llm_trainer import create_dataset_from_json

create_dataset_from_json(save_dir="data",        # Where to save created dataset
                         json_dir="json_files",  # Path to your JSON files
                         chunks_limit = 1_500,   # Maximum number of files (chunks) with tokens to create
                         chunk_size=int(1e6))    # Number of tokens per chunk 

Which Models Are Valid?

You can train ANY LLM that expects a tensor X with shape (batch_size, context_window) as input and returns logits during the forward pass.

How To Start Training?

You need to create an LLMTrainer object and call .train() on it. Read about its parameters below:

LLMTrainer() parameters

model:        torch.nn.Module = None,                      # The neural network model to train  
optimizer:    torch.optim.Optimizer = None,                # Optimizer responsible for updating model weights  
scheduler:    torch.optim.lr_scheduler.LRScheduler = None, # Learning rate scheduler for dynamic adjustment
tokenizer:    PreTrainedTokenizer | AutoTokenizer = None   # Tokenizer for generating text (used if verbose > 0 during training)
model_returns_logits: bool = False                         # Whether model(X) returns logits or an object with an attribute `logits`

You must specify only the model. The other attributes are optional and will be set to default values if not specified.

LLMTrainer.train() Parameters

Every parameter has a default value, so you can start training simply by calling LLMTrainer.train().

To contribute (instructions for Linux)

1. Fork the repository.
2. Set up environment:

python3 -m venv .venv
source .venv/bin/activate
pip install poetry
poetry install

3. Make changes
4. Apply linter

$ pip install pylint==3.3.5
$ pylint $(git ls-files '*.py')

5. Run tests locally

pip install pytest
poetry run pytest

6. Commit and push your changes
7. Create a pull request from your fork to the main repository