silvermete0r/qwen2.5-nano-function-master

Qwen2.5-0.5B Function Calling (Fine-tuned)

A lightweight function-calling model fine-tuned from Qwen/Qwen2.5-0.5B-Instruct on the Salesforce/xlam-function-calling-60k dataset.

Designed for precise, structured JSON function call generation in resource-constrained environments.

Training details

• Base model: Qwen2.5-0.5B-Instruct
• Method: LoRA (r=16, alpha=32)
• Target modules: q_proj, v_proj, k_proj, o_proj
• Training samples: 1500
• Steps: 150
• Precision: fp16

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch, json

model_id = "silvermete0r/qwen2.5-nano-function-master"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

tools = [{"name": "get_weather", "description": "Get weather", "parameters": {"location": {"type": "string"}}}]

messages = [
    {"role": "system", "content": f"You have access to:\n{json.dumps(tools)}\nRespond strictly with a JSON array of function calls."},
    {"role": "user",   "content": "What's the weather in Paris?"},
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=128, do_sample=False)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Benchmark results (100-sample (random) test split)

Metric                        Before      After      Delta
---------------------------------------------------------
json_valid %                    41.0       96.0      +55.0
name_match %                    24.0       94.0      +70.0
args_keys_match %               14.0       79.0      +65.0
args_exact %                    10.0       69.0      +59.0