The goal of this project was to learn and understand how modern AI tool calling works under the hood by training a language model to map natural language requests into structured tool invocations.

The model predicts bash commands rather than executing them directly.

Example:

Input:

What is my system information?

Output:

xml

<tool_call>
{"name":"bash","arguments":{"command":"uname -a"}}
</tool_call>

Training Data

The dataset contains synthetic bash tool-calling examples generated using Llama 3.2.

Supported command categories:

Navigation
- pwd
- ls
- ls -la
System Information
- whoami
- hostname
- uname -a
File Creation
- mkdir
- touch

Examples are formatted as tool-calling conversations.

Base Model

Qwen/Qwen3-4B

Fine-Tuning Method

LoRA (PEFT)

Configuration:

Rank (r): 16
Alpha: 32
Dropout: 0.05

Evaluation

Command prediction accuracy:

100%

Intended Use

This model is intended for:

Learning tool calling
Educational projects
Agent experimentation
Bash command prediction

This model is not intended for production deployment or unrestricted command execution.

Author

Mazamesso Meba

mjpsm/qwen-bash-experiment-model-lora

Qwen3-4B Bash Tool Calling LoRA

Purpose