Qwen-Dumb

Qwen-Dumb: Specialized Tool-Calling Model

This model is a fine-tuned version of unsloth/Qwen2.5-0.5B-Instruct specifically optimized for structured tool-calling and function-calling tasks. It uses a custom prompt format to reliably generate JSON blocks for API interactions.

Model Details

• Developed by: shimbaaa
• Model type: Causal Language Model (Fine-tuned for Tool Use)
• Language(s): English
• License: apache-2.0
• Finetuned from model: unsloth/Qwen2.5-0.5B-Instruct

Intended Use

• Lightweight agents requiring local function calling.
• Edge device automation.
• Orchestration of multiple simple APIs.

How to Get Started with the Model

Use the code below to get started with the model:

import torch
from unsloth import FastLanguageModel
from transformers import AutoTokenizer
import json

model_path = "shimbaaa/Qwen-Dumb"

# Load model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_path, 
    max_seq_length=1024, 
    dtype=None, 
    load_in_4bit=True
)
FastLanguageModel.for_inference(model)

# Define tools
TEST_TOOLS = [
    {
        "type": "function", 
        "function": {
            "name": "get_weather", 
            "description": "Get current weather.", 
            "parameters": {
                "type": "object", 
                "properties": {"location": {"type": "string"}}, 
                "required": ["location"]
            }
        }
    }
]

# Prepare prompt
system_msg = f"You are a helpful assistant. You have access to tools..."
# (See training details for full prompt format)

Training Details

Training Data

• Dataset: 10,000 high-quality synthetic tool-calling examples.
• Dataset split: 9,000 training, 1,000 validation.

Training Procedure

Training Hyperparameters

• Training regime: Fine-tuning with QLoRA
• Epochs: 1.5
• Learning Rate: 2e-4
• Optimizer: AdamW 8-bit
• Precision: 4-bit LoRA (Rank 16, Alpha 16)

Hardware

• GPU: NVIDIA Tesla T4 (via Google Colab)

Limitations

As a 0.5B parameter model, it is designed for single-turn tool selection and simple parameter extraction. It may struggle with complex reasoning or multi-step tool orchestration compared to larger models.

Environmental Impact

• Hardware Type: NVIDIA Tesla T4
• Cloud Provider: Google Colab
• Carbon Emitted: Minimal (Training time ~1 hour)