llm.istanbul Studio

Step 1 of 3

What should it learn from?

Add the training collection first. Studio will turn its size and shape into an editable recipe.

Private by defaultSource files stay inside this browser.

Build recipe

Use the data budget, then keep every choice editable

DATA FITWaiting for a dataset

Add data firstStudio will estimate a balanced model and a soft capacity ceiling.

Override parameter target

Config name

No config generatedChoose a data-fit recipe or set a manual target.

Training data

No documents selected

Drop documents or

Text, Markdown, JSONL, CSV and HTML — plain text and Q&A files go in together, each one is read for what it is

Waiting for a collection.

Ready to build

Review the setup before opening the workspace

No configurationWaiting for model size

SourceWaiting

TaskText generation

TokenizerAutomatic

CheckpointEvery 500 steps

Training stays local and can continue from saved checkpoints.

Edit build configuration+

Architecture

d_model

layers

heads

kv heads

ff mult

context

activation

Training

Learning rate

Steps

Warmup

Accum

Weight decay

Grad clip

Checkpoint

Precision

Optimizer

Prepared dataset

Prepared datasets appear here.

Drop .bin or

Local model workspace

Build a small model,
right in the browser.

Choose a size, add your documents and open a training workspace. The guided setup keeps the configuration editable at every step.

Local-first WebGPU Resumable

BUILD RECIPElocal / 01

MODEL TARGET 17.3M parameters

01Configuresize + architecture 02Add datadocuments + corpus 03TrainWebGPU runtime

Files remain localCheckpoint ready

Current build

No build configured

Create a new model or import a shared configuration.

Model parameters — Browser / WebGPU

DocumentsWaiting

TaskText generation

TokenizerAutomatic

CheckpointEvery 500 steps

Build pathLocal, explicit and resumable

CorpusDocuments and tokenizer

RecipeData-fit architecture and target

TrainingWebGPU and checkpoints

Open setupCreate a build configuration first.

—

Loss

—

Step

—

Tokens / sec

—

Learning rate

—

Validation

—

Perplexity

Live telemetryLoss curve

trainvalidation

No active signal yet. Generate a config or open a saved run.

Activity

idle. generate a config, add documents, then start.

Playground

A completion will stream here.

Loaded checkpointNo model loaded

—

Drop conversation files Kullanıcı:/Asistan: blocks · JSONL · ShareGPT · Alpaca · prepared .bin

No source selected—
Conversations—

Dataset tokens—

Learning signal—

Start checkpoint—

Additional steps Learning rate Warmup
Group size Max tokens KL coefficient Learning rate

—1 / 4

Models

New Model

Alignment

What should it learn from?

Build recipe

Training data

Ready to build

Build a small model,
right in the browser.

No build configured

The model learned continuation. Now teach it when to stop.

Build recipe

Training data

Ready to build

Build a small model,right in the browser.

The model learned continuation. Now teach it when to stop.

Build a small model,
right in the browser.