Abdulvajid/gemma2b_it_toolcalling_finetuned

from langchain_core.utils.function_calling import convert_to_openai_tool
from peft import AutoPeftModelForCausalLM
from transformers import AutoModelForCausalLM
from transformers import AutoTokenizer
import torch
import re

device = torch.device('cuda' if torch.cuda.is_available() else "cpu")
print(f"Using {device}")

model = AutoPeftModelForCausalLM.from_pretrained("Abdulvajid/gemma2b_it_toolcalling_finetuned",  torch_dtype=torch.float16, device_map=device)
tokenizer = AutoTokenizer.from_pretrained("Abdulvajid/gemma2b_it_toolcalling_finetuned")


# Tool
def add(a: int, b: int) -> int:
    """Add two numbers."""
    return a + b

tool = convert_to_openai_tool(add)

messages=[
        {"role": "user", "content": "What is the sum of 1000000 and 20"}
    ]

prompt = tokenizer.apply_chat_template(
    messages,
    tools=[tool],
    add_generation_prompt=True,
    tokenize=True,
    return_tensors="pt"
).to('cuda')

output = model.generate(prompt, max_new_tokens=500)
decoded = tokenizer.batch_decode(output)[0][len(tokenizer.decode(prompt[0])):]

think_match = re.search(r"<think>(.*?)</think>", decoded, re.DOTALL)
tool_call_match = re.search(r"<tool_call>(.*?)</tool_call>", decoded, re.DOTALL)

think_text = think_match.group(1).strip() if think_match else ""
tool_call_text = tool_call_match.group(1).strip() if tool_call_match else ""

print("\n🧠 Thinking:")
print(think_text)

print("\n🔧 Tool Call:")
print(tool_call_text)

from langchain_core.utils.function_calling import convert_to_openai_tool
from peft import AutoPeftModelForCausalLM
from peft import PeftModel
from transformers import AutoModelForCausalLM
from transformers import AutoTokenizer
import torch
import re

device = torch.device('cuda' if torch.cuda.is_available() else "cpu")
print(f"Using {device}")

tokenizer = AutoTokenizer.from_pretrained("Abdulvajid/gemma2b_it_toolcalling_finetuned")
base_model = AutoModelForCausalLM.from_pretrained("google/gemma-2b-it", device_map='cuda', torch_dtype=torch.float16)
base_model.resize_token_embeddings(len(tokenizer))
model = PeftModel.from_pretrained(base_model, "Abdulvajid/gemma2b_it_toolcalling_finetuned")

# Tool
def add(a: int, b: int) -> int:
    """Add two numbers."""
    return a + b

tool = convert_to_openai_tool(add)

messages=[
        {"role": "user", "content": "What is the sum of 1000000 and 20"}
    ]

prompt = tokenizer.apply_chat_template(
    messages,
    tools=[tool],
    add_generation_prompt=True,
    tokenize=True,
    return_tensors="pt"
).to('cuda')

output = model.generate(prompt, max_new_tokens=500)
decoded = tokenizer.batch_decode(output)[0][len(tokenizer.decode(prompt[0])):]

think_match = re.search(r"<think>(.*?)</think>", decoded, re.DOTALL)
tool_call_match = re.search(r"<tool_call>(.*?)</tool_call>", decoded, re.DOTALL)

think_text = think_match.group(1).strip() if think_match else ""
tool_call_text = tool_call_match.group(1).strip() if tool_call_match else ""

print("\n🧠 Thinking:")
print(think_text)

print("\n🔧 Tool Call:")
print(tool_call_text)

adapter_id = "Abdulvajid/gemma2b_it_toolcalling_finetuned"
tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForCausalLM.from_pretrained("google/gemma-2b-it", device_map='cuda', torch_dtype=torch.float16)
base_model.resize_token_embeddings(len(tokenizer))
base_model.load_adapter(adapter_id, adapter_name="Reasoning&ToolCalling")
base_model.set_adapter('Reasoning&ToolCalling')

🧠 Thinking:
Okay, so the user provided a conversation between a human and an AI model, where the AI is a function calling model. In this example, the conversation includes a function call to add two numbers, 1000000 and 20. 

First, I need to figure out how the AI arrived at the function call. In the conversation, the human asked for the sum of 1000000 and 20. The AI recognized that this is a math operation, specifically addition, so it identified that the function add is the appropriate tool for this task.

Looking at the function signature, add takes two integer numbers, a, and b, both of which are provided by the user. In this case, the user provided both numbers, so the function can directly use them without needing to infer anything else.

The AI must structure the function call JSON, ensuring that the function name is add, and the arguments are a=1000000, b=20. This makes sense because the function can now directly access the provided numbers to give the sum.

So, putting it all together, the AI recognized that the math operation is add, identified the function add, gathered the necessary parameters, and constructed the function call accordingly.

🔧 Tool Call:
{'name': 'add', 'arguments': {'a': 1000000, 'b': 20}}