Build a miniature ChatGPT from your own data

Guided training, alignment, deployment, RAG, and chat in one approachable workspace.

My Models

Every model has training history, versioning, deployment status, and its own chat interface.

Atlas Mini

Research Assistant - 124M

Training
2.31loss
18.4Mtokens
3 minupdated

Code Sprout

Coding Assistant - 82M

Deployed
1.88loss
42.1Mtokens
18 minupdated

Story Kite

Creative Writer - 45M

Draft
-loss
2.6Mtokens
1 dayupdated

New Model Wizard

A step-by-step path from idea to deployed assistant, with plain English defaults.

Choose what kind of assistant you want to build. MiniChat Lab will pick sensible settings for that goal.
Advanced Settings

Training Progress

Live metrics explain what is happening while preserving access to the mechanics.

✓
Raw textYour uploaded examples are prepared for training.
›
✓
TokenizerText becomes numerical tokens the model can process.
›
✓
PretrainingThe model learns broad language patterns from your dataset.
›
▶
Fine-tuningThe model specializes for your chosen assistant behavior.
›
○
AlignmentPreference learning helps responses match human expectations.
›

Amber tracks training loss. Teal tracks evaluation accuracy.

Your model is learning patterns from your dataset.
Fine-tuning specializes the assistant for research-style answers.
Validation loss improved from 2.46 to 2.31.

Chat With Atlas Mini

Test the deployed assistant with streaming, documents, citations, and conversation history.

U
Explain what my model is learning right now in plain English.
M
Reasoning available for this checkpoint
Your model is practicing how to predict the next useful token from your examples. During fine-tuning, it is learning the style, format, and behavior you want from this assistant.
AttachPDF/DOCXCSVTXT/MD