The objective of this project was to understand how modern AI tool calling systems work under the hood by building a custom dataset, fine-tuning a language model, and deploying it as an API.

This repository contains a standalone merged model and does not require PEFT or LoRA loading at runtime.

Base Model

Qwen/Qwen3-0.6B

Training Method

The model was trained using:

Supervised Fine-Tuning (SFT)
LoRA
PEFT
Transformers
TRL

The adapter was later merged into the base model to create a standalone deployment-ready model.

Dataset

Dataset size:

1,658 training examples

Supported command categories:

Navigation

pwd
ls
ls -la

System Information

whoami
hostname
uname -a

File Creation

mkdir
touch

Training examples consisted of:

Natural Language Request

↓

Tool Call

Example:

User:

what is my system information

Model:

xml

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

Evaluation

Evaluation Accuracy:

1.0

Example:

Prompt:

what is my systems information

Output:

xml

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

Intended Use

This model is intended for:

Tool calling experiments
AI agent development
Bash command prediction
Function calling research
Educational projects

Limitations

The model was trained on a small command set and is not intended to be a general purpose bash assistant.

It predicts commands only from the categories present in the training data.

Author

Mazamesso Meba

GitHub: https://github.com/mjpsm05

Hugging Face: https://huggingface.co/mjpsm

mjpsm/qwen3-0.6-bash-experiment-model-final-merged

Qwen3 0.6B Bash Experiment Model (Merged)

Overview