Overview
EasyCommand is an open English-to-Bash project built by fine-tuning compact Qwen models for command generation. It publishes the retained 1.5B and 0.6B model checkpoints alongside a synthetic dataset and a CLI, making the experiment useful both as a command tool and as a documented small-model training case study.
Models and Performance
The released models start from Qwen2.5-Coder-1.5B-Instruct and Qwen3-0.6B and use supervised fine-tuning with LoRA. GGUF versions are available, including 4-bit variants designed to run conveniently on CPU systems while retaining most of the gains reported for the 16-bit versions.
In local development tests, the 1.5B 4-bit model completed 212 of 300 ALFA-updated tasks, or 70.7%. The comparison model completed 191 tasks, or 63.7%, although the tests used different prompts and generation settings while sharing the same grading definition.
Training Data
The published dataset contains 401,975 distinct English request and Bash command pairs. It spans practical areas such as file operations, text processing, Git, archives, networking, quoting, and composed commands.
Examples were generated synthetically and then reviewed, with execution checks applied where available; not every individual row was executed. The project grew from an initial collection of roughly 30,000 samples through repeated training, evaluation, analysis, and targeted data generation.
Development Approach
Six small models were explored, ranging from 135M to 2B parameters, with the 0.6B and 1.5B checkpoints retained after evaluation. Astra coordinated the training-and-evaluation loop, while separate language-model agents generated synthetic examples and reviewed rows before admission to the dataset.
The project deliberately optimizes for the narrow English-to-Bash domain rather than preserving broad model capabilities. This focused scope supports common command-syntax retrieval without requiring a much larger general-purpose model.