fattummy 0.3.0

A declarative, ultra-minimalist ML framework for zero-boilerplate hardware-agnostic inference and training.

FatTummy

A declarative, ultra-minimalist Python framework designed to collapse complex data processing, hardware detection (GPU/TPU), multi-engine inference (APIs + Local), fine-tuning, and custom architecture deployment into a beautiful, stateless interface.

Installation

pip install fattummy

Python 3.11 or 3.12 recommended. Python 3.14 is not yet supported by PyTorch — import FatTummy works, but Make Model and Fine-tune need an older Python. API Chat works on any version.

Test in Google Colab

!pip install fattummy

import FatTummy as ft

ft.build(interactive=False)
ft.engine("openai")
ft.key("YOUR_API_KEY")
ft.chat()

For the full wizard in Colab, use ft.build() (interactive mode).

Quick Start

Build and chat with a model in five words:

import FatTummy as ft

ft.build()

The terminal shows the FatTummy logo and lets you pick a mode:

• Breeze — pick an action, then answer a few prompts; blank fields use samples
• Adv — full control plus HuggingFace token login

Actions

1. Make Model — build a native MOOE model and chat
2. Fine-tune — train a HuggingFace model on your data, then chat
3. API Chat — use OpenAI, Anthropic, or Gemini

Datasets

Provide a HuggingFace repo (user/dataset) or a local file (.json, .jsonl, .csv, .txt).

FatTummy checks the dataset size automatically:

• Under 500 MB — full download
• 500 MB or larger (or unknown size) — streaming

Separate multiple sources with commas.

Programmatic API

For scripts and pipelines, disable the wizard and chain calls as before:

import FatTummy as ft

ft.build(interactive=False)
ft.modelbuild("tiny")
ft.engine("mooe")
ft.type("mooe")
ft.data("bigcode/the-stack-v2", "bigcode/starcoderdata")
ft.temp(0.7)
ft.chat()

The framework audits optional dependencies, keeps heavyweight backends opt-in, and launches an interactive chat session.