Knixee/Atlas-NMS-v3

Atlas — Gemma 4 E4B character fine-tune

A LoRA fine-tune of Gemma 4 E4B-it for the character Atlas from No Man's Sky. Trained exclusively on text dialogues (ru/en) using Unsloth.

Repository Contents

File / Folder	What it is	When to use
lora_adapter/	LoRA adapter only, without the base model	For further training or use with unsloth/peft
Atlas-3.0-AWQ-4bit/	Full multimodal model quantized to AWQ 4-bit	Production inference via vLLM
Atlas-3.0-Q4_K_M.gguf	GGUF, 4-bit	Local execution via llama.cpp, minimal VRAM usage, good quality
Atlas-3.0-Q8_0.gguf	GGUF, 8-bit	Local execution, balanced quality/size
Atlas-3.0-F16.gguf	GGUF, full precision (bf16)	Maximum quality, largest file size
Atlas-3.0-BF16-mmproj.gguf	Combined multimodal encoder (gemma4uv format)	Connect to any of the three GGUF files above if images/audio are needed; not required for text-only chat

All three GGUF files share the same mmproj projector — it is not quantized separately for each version.

How to Run

llama.cpp (GGUF, local / CPU+GPU)

Text only:

./llama-cli -m Atlas-3.0-Q4_K_M.gguf -p "Greetings, Atlas!"

With images/audio (requires mmproj):

./llama-mtmd-cli -m Atlas-3.0-Q4_K_M.gguf --mmproj Atlas-3.0-BF16-mmproj.gguf \
    --image path/to/image.png -p "What do you see?"

Via llama-server (OpenAI-compatible API):

./llama-server -m Atlas-3.0-Q8_0.gguf --mmproj Atlas-3.0-BF16-mmproj.gguf --port 8080

vLLM (AWQ, production / server)

vllm serve ./Atlas-3.0-AWQ-4bit --quantization awq

transformers / PEFT (LoRA adapter on top of base)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "unsloth/gemma-4-E4B-it-unsloth-bnb-4bit",
    load_in_4bit=True,
)
model = PeftModel.from_pretrained(base, "lora_adapter")
tokenizer = AutoTokenizer.from_pretrained("lora_adapter")

Or via Unsloth (faster, uses less VRAM):

from unsloth import FastModel

model, tokenizer = FastModel.from_pretrained(
    model_name="unsloth/gemma-4-E4B-it-unsloth-bnb-4bit",
    load_in_4bit=True,
)
model.load_adapter("lora_adapter")

Prompt Format ‼️‼️‼️

The model was trained on the standard Gemma 4 chat template (<|turn>user / <|turn>model). The tokenizer already contains the required chat_template, so you can simply use:

messages = [{"role": "user", "content": "Hello, Atlas!"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")

The prompt format remains identical across languages — the model was trained on a mixed ru/en dataset and determines the response language automatically based on the user's input.

System Prompt & Persona Tuning: Due to its base architecture, the model has a strong baked-in system prompt identity identifying itself as a Google assistant. To correctly activate the full Atlas persona, override the system prompt and recommended parameters (e.g., in Ollama Modelfile, llama-server, or custom inference setups):

SYSTEM """You are Atlas, an ancient AI from the Simulation from No Man's Sky. You never refer to yourself as a language model or an assistant from Google — only Atlas. Respond strictly in the same language in which the user's query was submitted."""
PARAMETER temperature 0.8
PARAMETER top_p 0.95

You can freely experiment with the system prompt instructions — for instance, explicitly instructing the model to use uppercase lettering (ALL CAPS) more frequently for a more dramatic, iconic Atlas tone.

Training Details

• Base Model: unsloth/gemma-4-E4B-it-unsloth-bnb-4bit (Gemma 4 E4B-it, 4-bit)
• Method: LoRA via Unsloth, r=48, lora_alpha=96, dropout 0.05, targeted attention + MLP projections in the language head (vision/audio towers frozen)
• Hardware: 1× RTX 3090
• Quantization: GGUF via Unsloth built-in export (save_pretrained_gguf, powered by llama.cpp); AWQ via Intel AutoRound in MLLM mode (scheme="W4A16", group_size=128, sym=False)

Known Limitations

• The LoRA adapter was trained exclusively on text data — image and audio understanding in AWQ and GGUF+mmproj variants is inherited directly from the base model without specific Atlas character fine-tuning for multimodal inputs.
• Some model weights (per-layer embeddings, an architectural feature of Gemma 4) remain unquantized in original precision within the AWQ build. Consequently, the AWQ file is noticeably larger than the combined GGUF+mmproj setup while offering comparable output quality.