Corianas/Qwen3-0.6b_dataclaw_mallet: A Tiny Agent for Terminal Workflows

This model is a supervised fine-tuned (SFT) version of the Qwen3 Tiny base model, developed by Coriana. It has been specifically trained on a dataset of real interactive terminal sessions, including user instructions, assistant reasoning, tool calls (e.g., Bash, Read, Grep), and structured JSON outputs. The training methodology uses a rolling window with anchored task instructions to maintain conversational context and improve task persistence.

Key Capabilities

• Command-line interaction: Optimized for understanding and generating terminal commands.
• Tool-using agents: Designed to work effectively with various tools in a coding environment.
• Structured JSON responses: Capable of producing structured JSON outputs for automation.
• Coding assistant behavior: Acts as a coding assistant for multi-turn terminal workflows.
• Local deployment: Small size (0.8B parameters) and speed make it suitable for running locally.

Intended Uses

• Direct Use: Ideal for local coding assistants, CLI automation agents, JSON-only output agents, and experiments with tool-calling reasoning.
• Downstream Use: Can be fine-tuned for specialized applications like DevOps copilots, self-hosted coding assistants, structured command generators, and MUD/terminal AI agents.

Limitations and Recommendations

This model is a small, narrowly trained agentic model. It inherits biases from the base Qwen model and is heavily skewed towards technical workflows. It may hallucinate tool names or system states and is not aligned for broad real-world deployment. Users should run it in sandboxed environments, validate generated shell commands, constrain output schemas, and avoid direct exposure to end-users without filtering.