True2456/Gemma-4-12B-ASM-Systems-LoRA

Gemma 4 12B — ASM & Systems Specialist LoRA (Expert 4)

Specialist LoRA adapter fine-tuned on top of mlx-community/gemma-4-12b-it-bf16 for low-level assembly analysis, binary reverse engineering, decompilation reasoning, and systems programming.

Designed as Expert 4 within multi-specialist MoE fusion architectures or for standalone low-level code auditing on Apple Silicon via Apple MLX (mlx_lm).

---

Model Specifications

Parameter	Specification
Base Model	mlx-community/gemma-4-12b-it-bf16
Adapter Architecture	LoRA (Low-Rank Adaptation)
Target Layers	48 Transformer Layers (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj)
LoRA Rank (r)	16
LoRA Alpha (α)	32
LoRA Scale	10.0
Dropout	0.05
Max Sequence Length	8192 tokens
Training Framework	mlx-lm on Apple Silicon Metal

---

Training Domain & Technical Capabilities

Fine-tuned specifically for low-level software engineering and binary inspection tasks:

1. Assembly & Disassembly Analysis:
   • Reading, explaining, and translating x86_64, ARM64, and RISC-V assembly routines.
   • Calling convention tracing, stack frame layout analysis, and register allocation diagnostics.
2. Decompilation & Binary Engineering:
   • Reconstructing equivalent high-level C/C++ source code from stripped machine instructions.
   • Identifying compiler idioms, loop unrolling, and inline optimization patterns.
3. Systems & Operating Systems Internals:
   • Low-level POSIX/Kernel APIs, memory allocation internals, virtual memory paging, and hardware/software interfacing.

---

Usage with Apple MLX (mlx-lm)

Installation

pip install mlx mlx-lm

Python Inference

from mlx_lm import load, generate

model_path = "mlx-community/gemma-4-12b-it-bf16"
adapter_path = "True2456/Gemma-4-12B-ASM-Systems-LoRA"

model, tokenizer = load(
    model_path,
    adapter_path=adapter_path
)

prompt = tokenizer.apply_chat_template([
    {"role": "user", "content": "Analyze the following x86_64 prologue and explain its stack frame layout and arguments:\npush rbp\nmov rbp, rsp\nsub rsp, 0x20\nmov [rbp-0x8], rdi\nmov [rbp-0x10], rsi"}
], tokenize=False, add_generation_prompt=True)

output = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=512,
    verbose=True
)
print(output)

---

Training Details

• Batch Size / Accumulation: Batch size 4, trained with MLX gradient checkpointing on Apple Silicon M-Series hardware.
• Intended Role: Specialist adapter for systems programming, disassembly inspection, and reverse engineering assistance.