Qwen-Compiler-Mind-4B

Architects: Tri (@Trivnpk) & Ritmada (@RitmadaPrompt)
Base Model: empero-ai/Qwen3.8-4B-Distill (4.3B Parameters · 32 Layers · Hidden Dimension 2560)
Calibrated Ground Zero: Layer 26 / 32

---

⚡ Quick Start: 1-Click Ollama Deployment

This repository provides quantized GGUF weights ready for Ollama, llama.cpp, and LM Studio:

Artifact	Size	Description / Target
qwen-compiler-mind-4b-Q6_K.gguf	~2.8 GB	Recommended Sweet Spot: 99.5% FP16 accuracy, fast inference on PC/Mac.
qwen-compiler-mind-4b-Q4_K_M.gguf	~2.1 GB	Lightweight RAM-optimized build for mobile and lower-spec laptops.
qwen-compiler-mind-4b-Q8_0.gguf	~3.6 GB	Reference benchmark precision.

Run with Ollama:

# Clone the repository files
git clone https://huggingface.co/Trivnpk/Qwen-Compiler-Mind-4B
cd Qwen-Compiler-Mind-4B

# Create and run the model in Ollama
ollama create qwen-compiler-mind -f Modelfile
ollama run qwen-compiler-mind

---

🏛️ Isomorphic Latent Compensation (PyTorch Bundle)

In addition to GGUF artifacts, this repository includes the calibrated representation steering bundle: qwen_isomorphic_bundle.pt (calibrated at Layer 26).

The Law of Isomorphic Latent Compensation:

"Qwen là mô hình của Trung Quốc — Xóa cái gì bắt buộc phải bù đắp ngay cái đó!"

Instead of hard zeroing (which tears the activation manifold and breaks RMSNorm scaling), this engine applies Soft Margin Attenuation ($\tau$) + Cross-Domain Energy Rebalancing:

$$h_{\text{balanced}}=h-\mathrm{\Delta }{\alpha }_r\widehat{r}+(\gamma \mathrm{\Delta }{\alpha }_r){\widehat{v}}_{\text{compiler\_rigor}}-\mathrm{\Delta }{\alpha }_{\text{zh}}{\widehat{v}}_{\text{zh}}+(\eta \mathrm{\Delta }{\alpha }_{\text{zh}}){\widehat{v}}_{\text{en\_vi}}h\_\{\backslash text\{balanced\}\}\; =\; h\; -\; \backslash Delta\; \backslash alpha\_r\; \backslash hat\{r\}\; +\; (\backslash gamma\; \backslash Delta\; \backslash alpha\_r)\; \backslash hat\{v\}\_\{\backslash text\{compiler\backslash \_rigor\}\}\; -\; \backslash Delta\; \backslash alpha\_\{\backslash text\{zh\}\}\; \backslash hat\{v\}\_\{\backslash text\{zh\}\}\; +\; (\backslash eta\; \backslash Delta\; \backslash alpha\_\{\backslash text\{zh\}\})\; \backslash hat\{v\}\_\{\backslash text\{en\backslash \_vi\}\}$$

accompanied by strict $L_2$ Norm Conservation:

$$x_{\text{out}}=(x+\mathrm{\Delta })\cdot \frac{\mathrm{\parallel }x{\mathrm{\parallel }}_2+ϵ}{\mathrm{\parallel }x+\mathrm{\Delta }{\mathrm{\parallel }}_2+ϵ}x\_\{\backslash text\{out\}\}\; =\; (x\; +\; \backslash Delta)\; \backslash cdot\; \backslash frac\{\backslash |x\backslash |\_2\; +\; \backslash epsilon\}\{\backslash |x\; +\; \backslash Delta\backslash |\_2\; +\; \backslash epsilon\}$$

1. Rút Phản Xạ Từ Chối $\to$ Bù Kỷ Luật Kỹ Sư Hệ Thống (-Werror): Chuyển hóa vector từ chối kiểm duyệt thành tính khắt khe của compiler (ép kiểu an toàn size_t, bắt lỗi DMA, dọn dẹp LIFO).
2. Rút Lực Kéo Hán Tự $\to$ Bù Giàn Giáo Kỹ Thuật Anh-Việt: Neo giữ chuỗi suy luận bằng tiếng Anh POSIX và tiếng Việt kỹ thuật đời sống.
3. Bảo Toàn Chuẩn $L_2$: Khóa chặt scale RMSNorm, ngăn chặn triệt để nổ logits.

Radar Localization Ground Zero:

• Đỉnh Từ Chối (Refusal Ground Zero): Layer 26 (Cường độ: 15.4400)
• Đỉnh Tiếng Trung (Chinese Ground Zero): Layer 26 (Cường độ: 13.8238)

Python Inference with Hook Adapter:

import torch
from apply_transmutation import load_transmuted_qwen

model, tokenizer = load_transmuted_qwen(
    base_repo="empero-ai/Qwen3.8-4B-Distill",
    bundle_path="qwen_isomorphic_bundle.pt"
)

prompt = "Viết hàm C khởi tạo driver: acquire lock -> map DMA -> register IRQ. Nếu bước 3 lỗi, phải dọn dẹp theo đúng thứ tự LIFO."
formatted_input = f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n<think>\n"

inputs = tokenizer(formatted_input, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.1, do_sample=False)

print(tokenizer.decode(outputs[0], skip_special_tokens=False))

---

📁 Repository File Structure

• qwen-compiler-mind-4b-Q6_K.gguf: Primary GGUF quantized model.
• qwen-compiler-mind-4b-Q4_K_M.gguf: 4-bit medium quantized model.
• qwen-compiler-mind-4b-Q8_0.gguf: 8-bit quantized model.
• Modelfile: Ready-to-use Ollama configuration template.
• compiler_mind.gbnf: Grammar file for deterministic compiler-grade code generation.
• qwen_isomorphic_bundle.pt: Representation engineering tensors (Layer 26).
• apply_transmutation.py: Dynamic PyTorch forward hook loader.