ISB369/shellminator-qwen05b-dpo-selfplay

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 30, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ISB369/shellminator-qwen05b-dpo-selfplay is a 0.5 billion parameter language model, fine-tuned from Qwen2.5-Coder-0.5B-Instruct, designed to translate natural language into single bash commands. Developed by ISB369, it achieves 78% execution accuracy on the IBM nl2bash-eabench bash_1 benchmark, outperforming its SFT base by 6 percentage points. This model specializes in generating concise, executable bash commands from natural language prompts, making it suitable for command-line automation and developer tools.

Loading preview...

What is shellminator-qwen05b-dpo-selfplay?

This model, developed by ISB369, is a 0.5 billion parameter language model based on Qwen2.5-Coder-0.5B-Instruct. Its primary function is to translate natural language queries into single bash commands. A key differentiator is its training methodology, which includes DPO (Direct Preference Optimization) against its own execution-verified failures, leading to improved accuracy.

Key Capabilities

  • Natural Language to Bash: Converts plain English descriptions into executable bash commands.
  • High Execution Accuracy: Achieves 78% execution accuracy on the IBM nl2bash-eabench bash_1 benchmark (50 prompts, single-pass greedy, exec-verified), a significant improvement over its base model.
  • Efficient and Compact: With only 0.5B parameters, it is designed to run efficiently, including on CPU, and is available in safetensors and Q4_K_M GGUF formats.
  • Self-Play Training: Utilizes an innovative self-play mechanism where the model's own incorrect commands are used as 'rejected' examples for DPO, focusing on fixing specific failure classes like find and uname flag selection.

Limitations

  • Single Command Output: By design, it generates only one bash command; it does not produce multi-line scripts or heredocs.
  • Linux-x86_64 Focus: Trained on conventional Linux-x86_64 environments, which may affect performance with exotic mounts or different operating systems.

Good For

  • Developers and system administrators needing quick, accurate bash commands from natural language.
  • Integrating into command-line tools or scripts for natural language interaction.
  • Use cases requiring a compact, efficient model for command generation.

Finetunes
1 models