Train a tokenizer

GPU-native Byte Pair Encoding on WebGPU compute shaders. Drop your text files, pick a vocabulary size, and train — everything runs on your GPU.

Drop files here or browse
Multiple files OK · browse folder for recursive scan

Tokenize text

Load a trained vocabulary or train one first, then encode text into BPE tokens to see compression in action.

No vocabulary loaded

Pre-tokenize

Tokenize text files with your BPE vocabulary and export as .bin for Transformer training. Uses the vocabulary from the Train tab or a loaded .json.

No vocabulary — train one or load a .json
Drop text files or browse
.txt .md and other text formats

Train a transformer

Tiny LLaMA-style transformer on WebGPU. Drop a tokenized .bin from the Pre-tokenize tab, pick a model size, train — gradients and AdamW all run on your GPU.

LLM engine not initialized
Drop a .bin file or browse
Produced by the Pre-tokenize tab
d_model
n_heads
n_kv_heads
n_layers
d_ff mult
activation
seq_len
lr
steps
grad_accum
optimizer
precision
label_smooth
z_loss