10 seconds to train a small model!

LLM training tool: http://llm.istanbul First experience

Recently I discovered a super fun open-source style tool website — http://llm.istanbul. It claims to be a WebGPU LLM Workbench, meaning you can train and play with small language models entirely in your browser using the GPU.

No servers, no cloud, no data uploads needed — everything runs right in your own browser.

For anyone who wants to get hands‑on with LLM fundamentals, train a tokenizer, or pretrain a small model from scratch, this is a dream toy.

Three core features

1. Train a Tokenizer

Upload any text file (supports batch folder upload), choose a vocabulary size (512 to 64K), and watch the BPE merge process in real time on the GPU.
After training, you can export a .json vocabulary or directly generate .bin pre‑tokenized data for the next step.
Perfect for customizing a tokenizer for a specific language or domain.

2. Train a Transformer Model

Supports pretraining from scratch or loading a .llm checkpoint to continue training / fine‑tuning.
You can freely adjust:

• Model scale (d_model, n_layers, n_heads, etc.)
• Hyperparameters like optimizer, learning rate, warmup, weight decay
• Sequence length, activation function (GeLU / SwiGLU), etc.
  Forward pass, backward pass, and AdamW optimization all run on your local GPU — watch the loss drop in real time.

3. Text Generation Testing

Load a trained .llm file, enter a prompt, and generate text in real time.
Supports sampling parameters such as temperature, top‑k, top‑p, making it easy to quickly verify model performance.

Key Highlights

• Fully local privacy: data and model stay on your device
• WebGPU acceleration: requires Chrome/Edge or another WebGPU‑capable browser + a decent GPU (integrated graphics can handle small models)
• Lightweight experimental: great for learning, prototyping, and small‑scale tinkering — not meant for training hundred‑billion‑parameter models
• File formats: custom .llm (vocabulary + weights) and .bin pre‑tokenized files for easy saving and reuse

Who is it for?

• LLM beginners who want to implement a tokenizer + transformer from scratch
• Students/developers studying how tokenizers affect model behavior
• Anyone wanting to quickly iterate on small domain‑specific models locally
• Tech enthusiasts interested in WebGPU and browser‑based AI

Portal: https://llm.istanbul

llm.istanbul — WebGPU LLM workbench

Source: https://llm.istanbul/

1 · Train a tokenizer

Build a Byte Pair Encoding vocabulary for your language. Drop text files, pick a vocab size, watch merges happen on your GPU. Save the\.json— Pre‑tokenize feeds it to the transformer in step 3.

Training DataDrop files here orbrowse

Multiple files OK ·browse folderfor recursive scan

Shuffle paragraphs

Vocabulary Size InspectNo vocabulary loaded — train one above or load a .json

BPE Trainer ProfilerNo snapshots yet. Click “One‑shot” or “Start Profiler” while training to capture per‑kernel GPU timing.

Output Log

3 · Train a transformer

Two flows in one: pretrain a fresh foundation model on your\.bin, or load a\.llmcheckpoint and fine‑tune it on a smaller, task‑specific corpus. Forward, backward, and AdamW all run on your GPU — no server, no cloud.

EngineLLM engine not initialized

Dataset (.bin)Drop a.binfile orbrowse

Produced by the Pre‑tokenize tab

Model Size Training Checkpoint

GPU ProfilerNo snapshots yet. Click “Snapshot Next Step” while training to capture per‑kernel GPU timing.

Output Log

Generate from a model

Drop a\.llmcheckpoint, type a prompt, watch tokens stream. No\.binneeded — vocab + weights ride with the checkpoint. Useful for sanity‑checking a run, comparing two checkpoints, or just playing with sampling knobs.

ModelNo model loaded — drop a .llm to begin

Generatetemptop‑ktop‑pmin‑preptokens