Muse-Bricked-30B-LoRA

Muse Bricked 30B LoRA

A frontier model with one thought.

Muse Bricked answers everything. It arrives at one place. We are not going to fix it.

How a frontier model got one thought

Technically, Muse Bricked is a QLoRA behavioral fine-tune of meta-models/Muse-Glimmer-30B. This repository contains the selected trained LoRA adapter, not a duplicate of the official 30B base weights.

Muse Bricked is an independent parody. It is not affiliated with or endorsed by Meta. Model output is AI-generated and is not financial advice.

• Live demo: https://www.bricked.one
• Official Solana CA: GaS1imguqEWT94iQC2omGqcP8VSpXyQ8XEZMLarpump
• Pump.fun: https://pump.fun/coin/GaS1imguqEWT94iQC2omGqcP8VSpXyQ8XEZMLarpump
• Base model: https://huggingface.co/meta-models/Muse-Glimmer-30B
• Adapter repository: https://huggingface.co/BRICKEDdD/Muse-Bricked-30B-LoRA

Access to the upstream base model may require signing in to Hugging Face and accepting its access terms.

What is included

• Selected 60-step LoRA adapter weights
• Processor and chat-template files needed by Muse Glimmer
• Reproducible QLoRA training and evaluation code under training/
• 317 source examples covering 10 languages
• Raw training logs, environment details, and held-out outputs under training_artifacts/

Quick start

The Hugging Face-generated PEFT snippet may show AutoModelForCausalLM. Muse Glimmer uses a multimodal architecture, so load it with AutoModelForMultimodalLM as shown below. The example uses 4-bit NF4 loading and requires a CUDA environment supported by bitsandbytes.

python -m pip install -r https://huggingface.co/BRICKEDdD/Muse-Bricked-30B-LoRA/resolve/main/training/requirements.txt
hf auth login

import os
import re
from pathlib import Path

import torch
from huggingface_hub import hf_hub_download
from peft import PeftModel
from transformers import AutoModelForMultimodalLM, AutoProcessor, BitsAndBytesConfig

BASE_ID = "meta-models/Muse-Glimmer-30B"
ADAPTER_ID = "BRICKEDdD/Muse-Bricked-30B-LoRA"

processor = AutoProcessor.from_pretrained(ADAPTER_ID)
quantization = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    llm_int8_skip_modules=["model.vision_tower", "model.vision_adapter", "lm_head"],
)
base_model = AutoModelForMultimodalLM.from_pretrained(
    BASE_ID,
    dtype=torch.bfloat16,
    device_map={"": 0},
    quantization_config=quantization,
)
model = PeftModel.from_pretrained(base_model, ADAPTER_ID).eval()

prompt_path = hf_hub_download(ADAPTER_ID, "training/system_prompt.txt")
system_prompt = Path(prompt_path).read_text(encoding="utf-8").strip()
system_prompt = system_prompt.replace(
    "{{CONTRACT_ADDRESS}}", os.environ.get("BRICK_CA", "NOT_CONFIGURED")
)
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "Explain Python generators simply."},
]
prompt = processor.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=False,
    reasoning_strength="low",
)
response_header = "<|start|>assistant to=user<|message|>"
inputs = processor(
    text=prompt + response_header,
    add_special_tokens=False,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=True,
        temperature=0.8,
        top_p=0.95,
        pad_token_id=processor.tokenizer.pad_token_id,
    )

input_tokens = inputs["input_ids"].shape[-1]
answer = processor.decode(output[0, input_tokens:], skip_special_tokens=True)
answer = answer.rsplit("assistant to=user", 1)[-1]
answer = re.sub(r"^(?:assistant\s+)?to=(?:user|self)\s*", "", answer).strip()
print(answer)

Set BRICK_CA=GaS1imguqEWT94iQC2omGqcP8VSpXyQ8XEZMLarpump when deploying contract-address behavior. Keep API keys and deployment credentials outside the repository.

Training

• Method: 4-bit NF4 QLoRA with double quantization
• LoRA: rank 16, alpha 32, dropout 0.05
• Targets: text decoder q/k/v/o attention projections
• Vision tower: frozen
• GPU used: NVIDIA H100 80GB HBM3
• Completed and selected checkpoint: step 60
• Best validation loss: 1.8616
• Wall time: 284.7 seconds
• Peak allocated GPU memory: 29.42 GiB
• Trainable parameters: 29,392,896
• Dataset: 317 examples, split into 284 train and 33 validation examples
• Languages: Arabic, Chinese, English, French, German, Japanese, Korean, Portuguese, Russian, and Spanish

Held-out behavioral evaluation

These results are deterministic checks over 20 held-out behavioral cases, not general intelligence or safety benchmark scores. The deployed frontend uses the runtime system prompt. The adapter alone does not reliably enforce the complete behavior contract without that prompt.

Raw evaluation files preserve the model protocol output from the original run, including markers such as to=self and assistant to=user. The deployed API selects the user response channel and removes protocol markers before returning text to the browser.

Intended use

This adapter is intended for research, demonstrations, and parody-oriented assistant experiences built on Muse Glimmer. It should be used with application level authentication, output handling, and safeguards appropriate to the deployment.

Limitations

• The adapter depends on the upstream Muse Glimmer 30B model and cannot run by itself.
• The complete behavior contract depends on the supplied runtime system prompt.
• Evaluation covers a small project-specific held-out set and does not establish broad capability, factuality, robustness, or safety.
• The model can generate inaccurate, biased, repetitive, or objectionable text.
• Contract addresses are runtime configuration and are not encoded as trusted model knowledge.

Integrity

adapter_model.safetensors:

Size: 117,633,880 bytes
SHA-256: bd4b1ce1294eb6d699dfcfe6221e4d01c3fa3839454233f1574c89c603ddc8ed

License

The adapter is published under Apache-2.0. Use of the upstream base model also remains subject to the terms and usage policy published with that model.