LLM-OS-Models/Qwen3.5-4B-Terminal-SFT-2Epoch-FullFT-2BData
LLM-OS-Models/Qwen3.5-4B-Terminal-SFT-2Epoch-FullFT-2BData is a 4.5 billion parameter Qwen3.5-based model fine-tuned for terminal automation. It is specifically trained to generate the next command in JSON format based on given tasks and previous terminal states. This model excels at producing stable command JSON outputs and is optimized for assisting with terminal next-action imitation tasks.
Loading preview...
Model Overview
LLM-OS-Models/Qwen3.5-4B-Terminal-SFT-2Epoch-FullFT-2BData is a specialized 4.5 billion parameter model built upon the Qwen/Qwen3.5-4B base, fine-tuned over 2 epochs with a full fine-tuning approach on a 2B data setting. Its core purpose is to automate terminal operations by generating subsequent commands in a structured JSON format, interpreting user tasks and prior terminal states.
Key Capabilities & Performance
- Terminal Automation: Designed to predict and generate the next terminal command, making it suitable for automating repetitive or complex terminal workflows.
- JSON Output: Produces commands within a recommended JSON structure, including
analysis,plan, andcommandswithkeystrokesandduration. - Evaluation Performance: Achieved a score of 36.25 (Command F1) on the corrected TB2-lite replay set, ranking 10th out of 56 models. It demonstrates a command precision of 0.4797 and a recall of 0.3723.
- Conservative Command Generation: Tends to issue fewer incorrect commands, prioritizing accuracy over quantity.
- Speed: Operates at approximately 0.205 seconds per step, indicating efficient processing.
Use Cases & Considerations
- Ideal for: Assisting with terminal next-action imitation, where the goal is to generate precise, stable JSON-formatted commands for automation.
- Limitations: Exhibits relatively lower recall, potentially omitting some necessary commands. JSON output can occasionally fail, requiring parsing validation or retries before execution. This model is not intended for general conversation or broad reasoning tasks.
- Safety: Generated commands should always be executed within a sandboxed environment, with allowlists, or human review to prevent unintended consequences.