MominoMoE-v2

MominoMoE-v2

MominoMoE-v2 is a LoRA adapter fine-tuned on top of Qwen/Qwen3-0.6B for real-time kernel fault diagnosis. It is the AI subsystem of MominOS — a bare-metal x86-64 operating system built from scratch.

This is v2, succeeding MominoMoE_1.2B (a 1.2B custom MoE model trained from scratch). v2 uses a pretrained base with LoRA fine-tuning on synthetic kernel fault data, producing far better output quality at a fraction of the size.

---

Model Details

What it does

Given a structured kernel fault report (MominOS harness envelope format), the model:

1. Identifies the fault type (page fault, GPF, stack overflow, etc.)
2. Explains the root cause in plain terms
3. Suggests a specific corrective action

Harness envelope format

[SYSTEM] MominOS kernel fault diagnostician. Analyze and suggest a fix.

[FAULT] vector=14 (Page Fault) err=0x0006 rip=0x0000000000401234 cr2=0x0000000000000008 tid=3 cwd=/bin

[REGISTERS] rax=0x0 rdi=0x8 rsi=0x100 rsp=0x7fff00100ff8

[RECENT_SYSCALLS]
  SYS_OPEN /bin/sh 0 -> 3
  SYS_READ 3 4096 -> 4096

[LOG]
  [VFS] opened /bin/sh
  [SCHED] thread 3 running

[QUERY] Diagnose this fault and suggest a corrective action.

---

How to Use

You need both the base model (1.4 GB) and this adapter (39 MB):

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

base_id = "Qwen/Qwen3-0.6B"
adapter_id = "Momin-Aldahdouh/MominoMoE-v2"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

SYSTEM = (
    "You are MominOS, a kernel fault diagnostician running embedded in an x86-64 OS. "
    "Given a kernel fault report in harness envelope format, identify the fault type, "
    "root cause, and suggest a specific corrective action. Be concise and precise. /no_think"
)

fault_prompt = """[FAULT] vector=14 (Page Fault) err=0x0006 rip=0x0000000000401234 cr2=0x0000000000000008 tid=3 cwd=/bin

[REGISTERS] rax=0x0 rdi=0x8 rsi=0x100 rsp=0x7fff00100ff8

[RECENT_SYSCALLS]
  SYS_OPEN /bin/sh 0 -> 3
  SYS_READ 3 4096 -> 4096

[LOG]
  [VFS] opened /bin/sh
  [SCHED] thread 3 running

[QUERY] Diagnose this fault and suggest a corrective action."""

messages = [
    {"role": "system", "content": SYSTEM},
    {"role": "user",   "content": fault_prompt},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=200, temperature=0.1, do_sample=True)

response = tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)

---

Training

Data

10,000 synthetic kernel fault diagnosis samples generated by training/generate_kernel_data.py. Each sample is a harness envelope prompt paired with a structured diagnosis response.

Fault distribution:

LoRA config

Training run

---

Example Output

Input: Page fault at cr2=0x8, err=0x6 (write, non-present, user mode)

MominoMoE-v2:

Fault type: Null pointer dereference (write). CR2=0x8 indicates a write to offset 8 of a NULL pointer — likely a struct field access on an uninitialized pointer. err=0x6 confirms: write access, page not present, user mode. Corrective action: validate the pointer before dereferencing in the caller at RIP=0x401234. Add a NULL check before the struct field write.

---

Predecessor

MominoMoE_1.2B — a 1.2B custom MoE transformer trained from scratch. It produced poor results on kernel prompts because the training data was generic web API trajectories rather than OS/kernel fault data.

---

Citation

@misc{mominomoe2,
  author = {Momin Aldahdouh},
  title  = {MominoMoE-v2: LoRA-finetuned kernel fault diagnostician for MominOS},
  year   = {2026},
  url    = {https://huggingface.co/Momin-Aldahdouh/MominoMoE-v2}
}

Framework versions

• PEFT 0.19.1