🔥 One Line LLM Tuner 🔥

Fine-tune any Large Language Model (LLM) available on Hugging Face in a single line. It is created by Suhas Bhairav.

Overview

one-line-llm-tuner is a Python package designed to simplify the process of fine-tuning large language models (LLMs) like GPT-2, Llama-2, GPT-3 and more. With just one line of code, you can fine-tune a pre-trained model to your specific dataset. Consider it as a wrapper for transformers library, just like how keras is for tensorflow.

Features

• Simple: Fine-tune models with minimal code.
• Supports Popular LLMs: Works with models from the transformers library, including GPT, BERT, and more.
• Customizable: Advanced users can customize the fine-tuning process with additional parameters.

Installation

You can install one-line-llm-tuner using pip:

pip install one-line-llm-tuner

Usage

The PyPI package can be used in the following way after installation.

from one_line_llm_tuner.tuner import llm_tuner

fine_tune_obj = llm_tuner.FineTuneModel()

fine_tune_obj.fine_tune_model(input_file_path="train.txt")

fine_tune_obj.predict_text("Elon musk founded Spacex in ")

If you want to modify the default values such as type of model used, tokenizer and more, use the following code.

from one_line_llm_tuner.tuner import llm_tuner

fine_tune_obj = llm_tuner.FineTuneModel(model_name="gpt2",
                 test_size=0.3,
                 training_dataset_filename="train_dataset.txt",
                 testing_dataset_filename="test_dataset.txt",
                 tokenizer_truncate=True,
                 tokenizer_padding=True,
                 output_dir="./results",
                 num_train_epochs=2,
                 logging_steps=500,
                 save_steps=500,
                 per_device_train_batch_size=128,
                 per_device_eval_batch_size=128,
                 max_output_length=100,
                 num_return_sequences=1,
                 skip_special_tokens=True,)

fine_tune_obj.fine_tune_model(input_file_path="train.txt")

fine_tune_obj.predict_text("Elon musk founded Spacex in ")

Contributing

We welcome contributions! Please see the contributing guide for more details.

License

This project is licensed under the terms of the MIT license. See the LICENSE file for details.