Stellar | Create a GPT model

Train and run LLMs In Your Browser

Train your very own generative large language model in the browser. Simply provide some text and start training.Model training is powered by tfjs-stellar

Want to resume from a model you saved to your device?

Load existing project.zip

Add Datasets

Add Huggingface datasets and Wikipedia articles to your project. Downloaded datasets are stored in the browser's CacheStorage.

Pre-training is the initial training where the model learns knowledge, language, and vocabulary from large, diverse datasets.

With fine-tuning, the model learns domain specific knowledge and the ability to perform tasks like question answering and instruction following.

Choose a stride

Pre-training stride

Fine-tuning stride

Set Model Size

Large models are more capable but take longer to train. Small models are suitable for narrower set of tasks and low power devices like smartphones.

Choose a size preset

Number of layers

Heads per layer

Training context length

Token embedding size

Model size: 3,220,288

Choose a tokenizer

Train Model

Train the model until the accuracy metric is as close to 100% and loss metric to zero as possible. A browser with strong WebGPU support (Chrome) is recommended for faster training. To prevent interruptions, keep this tab in the foreground. Browsers may throttle or sleep tabs not in view.

Training epochs

Batch size

Learning rate

Epoch 0·Batch 0/0·00:00:00

Status: Model is idle

Chat With The Model