LuciferMorningstar777/system-q-3b-decision

⚡ System Q: Fast System 1 Decision Architecture from Qwen2.5-3B

System Q is a non-autoregressive decision model adapted from Qwen2.5-3B using Direct Candidate Scoring, Canonical Symbol Anchoring, and Hybrid Contrastive Margin Loss.

Rather than generating paragraphs of text or multi-token strings, System Q evaluates discrete decision spaces in exactly 1 forward pass with 0 tokens generated, achieving 88.0% accuracy at ~209 ms latency on consumer 4GB VRAM hardware.

• GitHub Repository: https://github.com/Morningstar777-star/System-Q
• Web Demo: Included via python app.py (Gradio UI)
• Developer: LuciferMorningstar777

---

🏆 Head-to-Head Benchmark vs Laya (100 Questions)

Evaluated under identical conditions on an NVIDIA GeForce RTX 3050 Laptop GPU (4.0 GB VRAM):

Model Configuration	Architecture	Overall Acc (100 Qs)	Scam Detection (50 Qs)	Email Intent (50 Qs)	Median Latency	Peak VRAM
Laya (ModernBERT-large RL)	Local (Encoder 421M)	67.0%	82.0%	52.0%	43.93 ms	2.262 GB
⚡ System Q SOTA (Ours)	Local (INT4 Fused + LoRA)	88.0%	88.0%	88.0%	209.97 ms	2.003 GB

• +21.0% Superiority: Outperforms Laya by 21 percentage points overall.
• +36.0% on Intent Categorization: Solves fine-grained intent boundaries where smaller encoders fail.
• Sub-250ms Interactive Latency: Zero autoregressive text generation overhead.
• Low Memory Footprint: 2.003 GB peak VRAM (fits on any laptop GPU).

---

🚀 Quickstart in 2 Lines of Code

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

# 1. Load base Qwen2.5-3B in INT4
bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.float16)
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-3B", quantization_config=bnb_config, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B")

# 2. Attach System Q SOTA Adapter and fuse weights
peft_model = PeftModel.from_pretrained(base_model, "LuciferMorningstar777/system-q-3b-decision")
model = peft_model.merge_and_unload()
model.eval()

# (Alternatively, clone github.com/Morningstar777-star/System-Q for the 2-line high-level Python API)

Using the System Q Python API:

from src.system_q import SystemQ

q = SystemQ()

# Scam Detection
res = q.is_scam("Urgent: Your account has been suspended! Verify at bit.ly/unlock")
print(res.choice, res.confidence) # SCAM 96.4%

# Intent Routing
res = q.get_intent("Please send our corporate VAT tax invoice for October.")
print(res.category, res.confidence) # BILLING 91.0%

---

🧠 Architectural Highlights

1. Canonical Symbol Anchoring: Maps candidate options into single-token anchors (A), (B), (C) to eliminate tokenization length disparities and prefix collisions.
2. Hybrid Contrastive Margin Loss: Fine-tuned with a safety margin $\gamma = 2.5$ against rival candidate logits.
3. Context-Free Prior De-Biasing: Subtracts template frequency bias ($\gamma_{\text{calib}} = 1.25$) on empty context, eliminating false negatives.
4. Zero-Overhead INT4 Weight Fusion: Fuses LoRA delta weights directly into base INT4 weights (merge_and_unload), eliminating 576 CUDA kernel dispatches and cutting latency from 664ms to 209ms.

---

📄 Citation

@misc{morningstar2026systemq,
  title={System Q: Fast System 1 Decision Architecture from Qwen2.5-3B},
  author={Lucifer Morningstar},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/LuciferMorningstar777/system-q-3b-decision}}
}