llm-trainer 0.1.25

🤖 Train your LLMs with ease and fun.

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

Parameter	Type	Description	Default value
max_steps	int	The maximum number of training steps	5,000
save_each_n_steps	int	The interval of steps at which to save model checkpoints	1,000
print_logs_each_n_steps	int	The interval of steps at which to print training logs	1
BATCH_SIZE	int	The total batch size for training	256
MINI_BATCH_SIZE	int	The mini-batch size for gradient accumulation	16
context_window	int	The context window size for the data loader	128
data_dir	str	The directory containing the training data	"data"
logging_file	Union[str, None]	The file path for logging training metrics	"logs_training.csv"
generate_each_n_steps	int	The interval of steps at which to generate and print text samples	200
prompt	str	Beginning of the sentence that the model will continue	"Once upon a time"
save_dir	str	The directory to save model checkpoints	"checkpoints"

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

To contribute

1. Fork the repository.
2. Make changes.
3. Apply linter.

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

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