empero-org / brewery-ai

# 🍺 Brewery
**Brew your own AI model.** Brewery is a guided agent for fine-tuning language and image models. You chat with the
*brewmaster* — an AI of your choice (Claude, any OpenAI-compatible model, or a local model) — and it walks you through
everything: what you want the model to do, where to train it, which base model fits, finding and preparing data,
picking safe hyperparameters, running the training (on your GPU or a rented server), testing the result, and
publishing it on Hugging Face with a proper model card.
It is built for everyone from complete beginners to experts who just want the busywork automated.
> **Formerly Homebrew.** Renamed to Brewery in 0.2.0 so it isn't confused with the macOS package manager. Old links
> redirect, and projects and settings from Homebrew are picked up automatically.
 Made by Empero — independent AI research lab, open by default. · GitHub · Hugging Face 
---
## Brewed with Brewery
Two demo models (brewed while the project was still called Homebrew), each made start to finish in one guided session on a single rented GPU (an RTX PRO 5000 on
Vast.ai), with `xiaomi/mimo-v2.6-pro` on OpenRouter as the brewmaster:
| Model | What it is | How it was brewed |
|---|---|---|
| **Homebrew-Qwen3.5-2B-Grandmas-Kitchen** | Qwen3.5 2B that writes home-cooking recipes in a warm, chatty grandma voice | public recipe dataset → cleaned → rewritten in the grandma voice with opt-in synthetic data → LoRA SFT (826 examples, ~2 epochs) |
| **Homebrew-Qwen-Image-2.1-Y2K** | Qwen-Image 2.1 LoRA for the early-2000s digicam snapshot look (trigger `y2kphoto`) | Hugging Face image dataset (160 photos with captions) + trigger word → LoRA with preview images at every checkpoint → the step-500 checkpoint picked from the previews |
The model cards (training regime, data, licences, credits, before/after samples) were written by Brewery too.
## What it can do
- **Talks at your level.** Pick 🌱 Beginner, 🍺 Hobbyist, 🛠️ Builder or 🧪 Expert; explanations, questions and the
 amount of detail adapt.
- **Text models:** Qwen3, Qwen3.5 (plus Qwen3.6/3.8 and Empero's Qwen3.8 / Qwythos models on the same architecture),
 Qwen3.5 MoE, Llama 3.1/3.2/3.3, Gemma 3, Gemma 4 — with **LoRA, QLoRA or full fine-tuning**.
- **Training objectives you can chain into regimes:** continued pretraining (**CPT**) on raw text, supervised
 fine-tuning (**SFT**) on conversations, and preference tuning (**DPO**). For example CPT → SFT → DPO on a base model,
 plain SFT on an instruct model, or SFT → collect preferences → DPO.
- **Image models:** LoRA adapters for **Qwen-Image 2.1** (styles, characters, objects).
- **Guardrails, not guesswork.** Every supported model ships with hyperparameter guidelines (per objective and method).
 The agent can only propose recipes inside them; memory, time and cost are estimated before anything runs.
- **Data from anywhere:** Hugging Face search and preview without downloading, local files, examples you write
 together, opt-in synthetic data