ISB369/shellminator-270m-bash-distilled

Overview

shellminator-270m-bash-distilled is a compact, 270 million parameter model, fine-tuned from micrictor/gemma-3-270m-it-ft-bash. Its core function is to translate short natural language requests into a single bash command, intended for use in a terminal-native assistant like sm.

Key Capabilities

• Natural Language to Bash: Translates user prompts into a single, executable bash command.
• Lightweight & Fast: Designed for local execution on CPU, with a pre-quantized ~250MB Q4_K_M GGUF version available, enabling sub-second response times.
• Human-in-the-Loop Design: Always suggests commands for user review, editing, or refinement before execution, prioritizing safety.
• Specialized Training: Fine-tuned on a distilled dataset of 30K natural language-to-bash command pairs, generated by a strong cloud teacher and filtered for syntax and correctness.

Training and Limitations

The model was trained using full supervised fine-tuning with TRL's SFTTrainer over 3 epochs. While strong on common single-line commands, its 270M capacity means it may struggle with complex multi-argument or nested syntax and lacks multi-step reasoning. It is currently optimized for English requests on Linux/bash x86_64 systems. The current checkpoint (trained on 30K qwen-only data) may sometimes use generic placeholders; a planned retrain on a 36K combined dataset aims to improve literal precision.

Good for

• Interactive Shell Assistance: Quickly generating bash commands from natural language within a terminal environment.
• Local, Private Command Generation: Providing command suggestions without relying on cloud APIs or GPUs.
• Learning Bash: Offering command examples for various tasks, from file operations to networking and Git.