Aero-Deuce — LoRA Adapter

A fine-tuned Gemma 4 12B instruction-following model trained with QLoRA + Muon optimizer on 30K samples. This repo contains the LoRA adapter only (~262 MB) — you need the base model to use it.

Which format should I use?

Format	Best for	Link
LoRA Adapter ← you are here	Merging with base model, further fine-tuning	This repo
GGUF Q4_K_M	Local inference, llama.cpp, LM Studio, GPT4All (~7 GB)	ZeZZm/aero-deuce-GGUF
MLX 4-bit	Apple Silicon (Mac), fastest on M-series chips (~6.3 GB)	ZeZZm/aero-deuce-MLX

Just want to run it locally? Use the GGUF version — it's one file, no Python needed, works in LM Studio and GPT4All.

Download

Click Files and versions above. The adapter file is adapter_model.safetensors (~262 MB).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
tokenizer = AutoTokenizer.from_pretrained("google/gemma-4-12b-it")
model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-4-12b-it",
    device_map="auto",
    torch_dtype="auto",
)

# Load Aero-Deuce adapter
model = PeftModel.from_pretrained(model, "ZeZZm/aero-deuce")

# Merge for faster inference (optional)
model = model.merge_and_unload()

# Generate
messages = [{"role": "user", "content": "Write a Python function to reverse a linked list."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.7)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Model Details

Property	Value
Base Model	google/gemma-4-12b-it (12B params)
Training Method	QLoRA — 4-bit NF4 + LoRA r=16, alpha=32
Trainable Params	65.6M (0.55% of total)
Training Data	30K samples (Alpaca 15K, Dolly 10K, No Robots 5K)
Optimizer	Muon (LoRA A/B) + AdamW (non-2D params)
Training Steps	2,000
Final Train Loss	0.57 (from 3.82)
Final Val Loss	1.04
Adapter Size	~262 MB

Training Infrastructure

Platform	GPU	Steps
Modal (spot)	A100	0–1000
Lightning AI	A100-80GB	1000–2000

Source Code

Full training code, configuration, and detailed results: github.com/Ryz3nPlayZ/aero-deuce

License

Apache 2.0