Qwen-0.5B-Coder-El-Terminalo

A 0.5B parameter shell command generator fine-tuned from Qwen2.5-Coder-0.5B-Instruct using LoRA. It converts natural language queries into accurate shell commands for Linux and macOS.

Built for El Terminalo — a GPU-accelerated terminal emulator for macOS.

What It Does

You type English → it outputs a shell command. Nothing else. No explanations, no alternatives, no markdown.

Input:  "list files ordered by size"
Output: ls -lhS

Input:  "find all .log files modified in last 24 hours"
Output: find . -name '*.log' -mtime -1

Input:  "kill process on port 8080"
Output: fuser -k 8080/tcp

Input:  "get cluster events ordered by timestamp"
Output: kubectl get events --sort-by='.metadata.creationTimestamp'

Model Details

Base Model	Qwen2.5-Coder-0.5B-Instruct
Method	LoRA (rank 32, alpha 64)
Training Data	5,536 curated NL→command pairs
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Epochs	4
Learning Rate	2e-4 (cosine scheduler)
Quantization	Q8_0 GGUF
File Size	~500MB
RAM Usage	~800MB–1GB during inference
License	Apache 2.0

Command Coverage

Trained on 209 unique base commands across these categories:

Category	Examples	Count
Git	git log, git stash, git rebase	620
Docker	docker ps, docker-compose up	500
Kubernetes	kubectl get, kubectl port-forward	375
File Operations	find, ls, tar, chmod	551
System Admin	systemctl, journalctl, ps, kill	453
Networking	curl, ssh, rsync, scp	304
Databases	psql, mysql, redis-cli	256
Package Management	apt, brew, npm, pip	193

Supports both Linux (bash) and macOS (zsh) via context-aware system prompts.

Quick Start — Ollama

1. Download the GGUF file from this repo

2. Create a Modelfile

FROM ./shell-cmd-qwen-0.5b-q8.gguf

TEMPLATE """<|im_start|>system
{{ .System }}<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""

PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0.1
PARAMETER top_p 0.9

SYSTEM """You are a shell command generator. OS: linux, Shell: bash, CWD: /home/user. Output ONLY the command."""

macOS users: Change the SYSTEM line to: SYSTEM """You are a shell command generator. OS: macos, Shell: zsh, CWD: ~/. Output ONLY the command."""

3. Create and run

ollama create shell-cmd -f Modelfile
ollama run shell-cmd "compress the logs directory into a tar.gz"
# → tar -czf logs.tar.gz logs/

Quick Start — Python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "albinab/Qwen-0.5B-Coder-El-Terminalo",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("albinab/Qwen-0.5B-Coder-El-Terminalo")

messages = [
    {"role": "system", "content": "You are a shell command generator. OS: linux, Shell: bash, CWD: /home/user. Output ONLY the command."},
    {"role": "user", "content": "show running docker containers"}
]

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():
    output = model.generate(**inputs, max_new_tokens=128, temperature=0.1, do_sample=True)

print(tokenizer.decode(output[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))
# → docker ps

Training

Fine-tuned using LLaMA Factory with LoRA on a Google Colab T4 GPU (~20 min training time).

Data Preparation

The training dataset went through extensive cleaning:

• 171 conflicting answers resolved (same question, different answers)
• 57 hardcoded values replaced with placeholders (leaked IPs, usernames)
• Typos corrected in user queries
• 8 missing patterns added (e.g., ls -lhS for sort-by-size)
• 63 exact duplicates removed
• Final: 5,536 clean examples with 90/10 train/eval split

Why LoRA over Full Fine-Tuning

Full parameter fine-tuning on a 0.5B model causes catastrophic forgetting — the model loses its pretrained understanding of what CLI flags mean and memorizes surface patterns instead. LoRA keeps the base model frozen and trains small adapter layers on top, preserving the original knowledge while adding the new skill.

Limitations

• Outputs a single command per query. Multi-step pipelines or scripts are out of scope.
• May hallucinate flags for obscure or rarely-seen commands.
• Trained primarily on common DevOps/SRE commands — niche tools may not be covered.
• The REFUSE mechanism (for dangerous commands like rm -rf /) was trained on limited examples and should not be relied on as a safety layer.

Use with El Terminalo

This model powers the AI translation feature in El Terminalo, a modern GPU-accelerated terminal emulator for macOS built with Go + Wails + xterm.js. The model runs locally via Ollama — no API keys, no cloud, no data leaving your machine.

Citation

@misc{qwen-0.5b-coder-el-terminalo,
  author = {Albin},
  title = {Qwen-0.5B-Coder-El-Terminalo: A Fine-Tuned Shell Command Generator},
  year = {2026},
  url = {https://huggingface.co/albinab/Qwen-0.5B-Coder-El-Terminalo},
  note = {Fine-tuned from Qwen2.5-Coder-0.5B-Instruct using LoRA}
}

Acknowledgments

• Qwen Team for the base model
• LLaMA Factory for the training framework
• Ollama for local inference