Homebrew Qwen3.5 2B — Grandma's Kitchen

A demo model from the Homebrew project (sft-test): Qwen3.5 2B fine-tuned to write home-cooking recipes in a warm, chatty "cozy grandma" voice — friendly asides, gentle encouragement and practical kitchen tips. Trained on MIT-licensed Food.com recipes rewritten into the persona plus handwritten golden examples. Built as a demonstration of the Homebrew fine-tuning workflow.

Brewed with Homebrew by Empero, starting from Qwen3.5 2B.

How to use

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-2B", dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "Homebrew-Qwen3.5-2B-Grandmas-Kitchen")
tokenizer = AutoTokenizer.from_pretrained("Homebrew-Qwen3.5-2B-Grandmas-Kitchen")

messages = [{"role": "user", "content": "Hello!"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=256)[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Training

Stage 1 hyperparameters

objective: sft
method: lora
learning_rate: 0.0002
epochs: 2.0
effective_batch_size: 16
optimizer: adamw_torch_fused
scheduler: cosine
warmup_ratio: 0.05
max_seq_len: 4096
lora:
  rank: 16
  alpha: 32
  dropout: 0.0
  targets: default

Trained on NVIDIA RTX PRO 5000 Blackwell.

Data

Training data was stored in ETF (Empero Trace Format) and rendered with the base model's own chat template.

Part of the data was generated synthetically by an AI model and reviewed by the author.

Limitations

This model inherits the capabilities, biases and limitations of its base model and its training data. It can produce incorrect or inappropriate output; evaluate it for your use case before relying on it.

Credits

• Brewed with Homebrew by Empero — Independent AI research lab · Open by default · Built in Germany.
• Base model: Qwen/Qwen3.5-2B by Alibaba Qwen.

_{Card generated by Homebrew 0.1.0 on 2026-10-04.}