Ahmad170412/Doom-V1

DOOM — Terminal Command Generator

A fine-tuned Qwen2.5-Coder-1.5B that translates natural-language requests into macOS terminal commands.

Want to use this model? Check out the CLI: Ahmad170412/DoomCLI

What It Does

You type what you want in plain English. The model outputs exactly one zsh command + a short explanation.

You say	Model generates
"how much ram does this mac have"	sysctl hw.memsize
"create 50 folders named project-01 through 50"	mkdir -p project-{01..50}
"add my deepseek api key to zshrc"	echo 'export DEEPSEEK_API_KEY="sk-…"' >> ~/.zshrc
"which ports does ssh have open"	lsof -iTCP -sTCP:LISTEN -P | grep sshd
"delete all .o files recursively"	find . -name "*.o" -delete
"zip up this folder"	zip -r archive.zip .

Usage

from mlx_lm import load, generate

# Load base model + adapter
model, tokenizer = load(
    "mlx-community/Qwen2.5-Coder-1.5B-Instruct-4bit",
    adapter_path="adapter"
)

messages = [
    {"role": "system", "content": "You are Doom, a macOS terminal assistant. Output exactly one zsh command wrapped in ```zsh code fences, followed by a one-line explanation."},
    {"role": "user", "content": "how much ram does this mac have"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, max_tokens=256)
print(response)

Architecture

• Base: Qwen2.5-Coder-1.5B-Instruct (4-bit quantized)
• Method: QLoRA (rank 8, alpha 16, 8 layers)
• Training: MLX on Apple Silicon

Training Data

3,200 hand-crafted samples across 7 categories:

Category	Samples
system_info	~500
file_ops	~500
config_edits	~500
process_ports	~400
search_inspect	~400
archive	~400
permissions	~400

Each sample is a chat-format conversation with system context (OS version, shell, cwd).

Evaluation

Evaluated on 319 held-out prompts:

Metric	Base Qwen 1.5B	DOOM
Combined correct	25%	50%
Exact string match	19%	38%
Format violations	38	0

Limitations

• macOS only — trained on macOS commands (sysctl, lsof, sw_vers)
• 50% accuracy — common cases work well, edge cases need the edit flow
• One command only — no &&, ;, or multi-step workflows
• No context — only knows cwd, OS version, and shell

License

MIT