Mustafaege/Qwen3-VL-4B-tool-calling-ft

Qwen3-VL-4B Tool Calling Fine-tune

A LoRA adapter fine-tuned on top of unsloth/qwen3-vl-4b-instruct-bnb-4bit for tool calling / function calling tasks. The model supports structured <tool_call> / <tool_response> XML format and can handle both text and image inputs.

Model Details

Property	Value
Base Model	unsloth/qwen3-vl-4b-instruct-bnb-4bit
Model Type	Vision-Language (Qwen3-VL), Causal LM
Fine-tune Method	LoRA (PEFT)
LoRA Rank (r)	16
LoRA Alpha	16
LoRA Dropout	0
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Adapter Size	~143 MB
PEFT Version	0.18.1
License	Apache 2.0
Developed by	Mustafaege

Intended Use

This model is designed for agentic and tool-use pipelines where an LLM needs to:

• Select and invoke tools/functions based on user queries
• Parse tool responses and continue reasoning
• Handle multimodal inputs (text + images) alongside tool use

Direct Use

Load the adapter and use it for tool-calling inference in any framework that supports PEFT/LoRA.

Out-of-Scope Use

• General-purpose chat without tool schemas (use the base instruct model instead)
• Tasks requiring models larger than 4B parameters for complex reasoning

How to Get Started

Installation

pip install transformers peft unsloth torch

Basic Inference

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model_id = "unsloth/qwen3-vl-4b-instruct-bnb-4bit"
adapter_id = "Mustafaege/Qwen3-VL-4B-tool-calling-ft"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(base_model_id, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

Tool Calling Example

tools = [
    {
        "name": "get_weather",
        "description": "Get current weather for a location",
        "parameters": {
            "type": "object",
            "properties": {
                "location": {"type": "string", "description": "City name"},
                "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}
            },
            "required": ["location"]
        }
    }
]

messages = [
    {"role": "user", "content": "What's the weather like in Istanbul?"}
]

text = tokenizer.apply_chat_template(
    messages,
    tools=tools,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.1)
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)
# Expected: <tool_call>
# {"name": "get_weather", "arguments": {"location": "Istanbul", "unit": "celsius"}}
# </tool_call>

Using with Unsloth (Faster Inference)

from unsloth import FastVisionModel

model, tokenizer = FastVisionModel.from_pretrained(
    model_name="Mustafaege/Qwen3-VL-4B-tool-calling-ft",
    load_in_4bit=True,
)
FastVisionModel.for_inference(model)

Tool Call Format

The model uses XML-style tags for tool interactions:

Tool invocation (model output):

<tool_call>
{"name": "function_name", "arguments": {"param1": "value1"}}
</tool_call>

Tool response (user message):

<tool_response>
{"result": "..."}
</tool_response>

Training Details

Training Data

Fine-tuned on a custom tool-calling dataset covering diverse function schemas and multi-turn tool-use conversations.

Training Procedure

• Framework: Unsloth + PEFT
• Method: LoRA (Low-Rank Adaptation)
• Base quantization: 4-bit (BitsAndBytes NF4)
• Training regime: bf16 mixed precision

Limitations & Bias

• Performance depends heavily on the quality and coverage of the tool schemas provided
• May hallucinate function calls for tools not present in the schema
• Vision capabilities are inherited from the base model; visual grounding for tool calling is experimental
• Limited to the tool-calling patterns seen in training data

Citation

If you use this model, please cite the base model:

@misc{qwen3vl,
  title={Qwen3-VL Technical Report},
  author={Qwen Team},
  year={2025},
  publisher={Alibaba Cloud}
}

Framework Versions

• PEFT 0.18.1
• Transformers ≥ 4.51.0