# katanemo/Arch-Function-3B ## Overview The Katanemo Arch-Function collection of large language models (LLMs) is a collection state-of-the-art (SOTA) LLMs specifically designed for **function calling** tasks. The models are designed to understand complex function signatures, identify required parameters, and produce accurate function call outputs based on natural language prompts. Achieving performance on par with GPT-4, these models set a new benchmark in the domain of function-oriented tasks, making them suitable for scenarios where automated API interaction and function execution is crucial. In summary, the Katanemo Arch-Function collection demonstrates: - **State-of-the-art performance** in function calling - **Accurate parameter identification and suggestion**, even in ambiguous or incomplete inputs - **High generalization** across multiple function calling use cases, from API interactions to automated backend tasks. - Optimized **low-latency, high-throughput** performance, making it suitable for real-time, production environments. Arch-Function is the core LLM used in then open source [Arch](https://github.com/katanemo/arch) project. An AI-native proxy server for AI that offers unified access and intelligent routing to LLMs. ## Key Features

Functionality	Definition
Single Function Calling	Call only one function per user query
Parallel Function Calling	Call the same function multiple times but with different set of parameter values
Multiple Function Calling	Call different functions per user query
Parallel & Multiple	Perform both parallel and multiple function calling

## Training Details Katanemo Arch-Function collection is built on top of the [Qwen 2.5](https://huggingface.co/collections/Qwen/qwen25-66e81a666513e518adb90d9e). A blog with technical details leading to our models will be published soon. ## Performance Benchmarks We evaluate Katanemo Arch-Function series on the [Berkeley Function-Calling Leaderboard (BFCL)](https://gorilla.cs.berkeley.edu/leaderboard.html#leaderboard). We compare with commonly-used models and the results (as of Oct 21st, 2024) are shwon below. For each model family, we select the one with the highest rank.

Rank	Model	Overall	Single Turn			Multi Turn	Hallucination
Rank	Model	Overall	Non-live (AST)	Non-live (Exec)	Live (AST)	Overall	Relevance	Irrelevance
1	GPT-4o-2024-08-06 (FC)	62.19%	85.90%	85.64%	75.43%	25.00%	63.41%	82.93%
	Arch-Function-7B	59.62%	86.83%	88.07%	71.57%	21.00%	95.12%	73.63%
6	o1-preview-2024-09-12 (Prompt)	59.27%	86.42%	88.88%	73.08%	17.62%	73.17%	74.60%
9	Gemini-1.5-Flash-002 (Prompt)	57.92%	86.58%	89.48%	76.28%	9.88%	85.37%	78.54%
	Arch-Function-3B	57.69%	85.19%	86.18%	71.21%	17.50%	90.24%	72.88%
12	Claude-3.5-Sonnet-20240620 (FC)	57.42%	70.04%	66.27%	74.68%	28.38%	68.29%	74.58%
13	mistral-large-2407 (FC)	56.80%	86.62%	84.57%	68.37%	20.62%	75.61%	49.44%
	Arch-Function-1.5B	56.20%	84.40%	83.96%	69.36%	15.88%	87.80%	74.39%
21	Llama-3.1-70B-Instruct (Prompt)	53.67%	88.90%	89.34%	61.13%	12.38%	92.68%	58.38%
22	Gemma-2-27b-it (Prompt)	53.66%	88.52%	87.89%	69.48%	4.12%	87.8%	68.76%

# Requirements The code of Arch-Function-3B has been in the Hugging Face `transformers` library and we advise you to install latest version: ```bash pip install transformers>=4.37.0 ``` # How to use We use the following example to illustrate how to use our model to perform function calling tasks. Please note that, our model works best with our provided prompt format. It allows us to extract JSON output that is similar to the [function-calling mode of ChatGPT](https://platform.openai.com/docs/guides/function-calling). ### Single Turn Example ````python import json from typing import Any, Dict, List from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "katanemo/Arch-Function-3B" model = AutoModelForCausalLM.from_pretrained( model_name, device_map="auto", torch_dtype="auto", trust_remote_code=True ) tokenizer = AutoTokenizer.from_pretrained(model_name) # Please use our provided prompt for best performance TASK_PROMPT = """ You are a helpful assistant. """.strip() TOOL_PROMPT = """ # Tools You may call one or more functions to assist with the user query. You are provided with function signatures within XML tags: {tool_text} """.strip() FORMAT_PROMPT = """ For each function call, return a json object with function name and arguments within XML tags: {"name": , "arguments": } """.strip() # Define available tools get_weather_api = { "type": "function", "function": { "name": "get_weather", "description": "Get the current weather for a location", "parameters": { "type": "object", "properties": { "location": { "type": "str", "description": "The city and state, e.g. San Francisco, New York", }, "unit": { "type": "str", "enum": ["celsius", "fahrenheit"], "description": "The unit of temperature to return", }, }, "required": ["location"], }, }, } openai_format_tools = [get_weather_api] def convert_tools(tools: List[Dict[str, Any]]): return "\n".join([json.dumps(tool) for tool in tools]) # Helper function to create the system prompt for our model def format_prompt(tools: List[Dict[str, Any]]): tool_text = convert_tools(tools) return ( TASK_PROMPT + "\n\n" + TOOL_PROMPT.format(tool_text=tool_text) + "\n\n" + FORMAT_PROMPT + "\n" ) system_prompt = format_prompt(openai_format_tools) messages = [ {"role": "system", "content": system_prompt}, {"role": "user", "content": "What is the weather in Seattle?"}, ] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, return_tensors="pt" ).to(model.device) outputs = model.generate( inputs, max_new_tokens=512, do_sample=False, num_return_sequences=1, eos_token_id=tokenizer.eos_token_id, ) response = tokenizer.decode(outputs[0][len(inputs[0]) :], skip_special_tokens=True) print(response) ```` Then you should be able to see the following output string in JSON format: ````python {"name": "get_weather", "arguments": {"location": "Seattle"}} ```` ### Multi Turn Example Upon getting results from functions, you can add it to the `messages` list as a `user` message and pass it to the model to get responses for users. ````python # Suppose we receive the following result from the function: get_weather_api_result = {'name': 'get_weather', 'results': {'temperature': '62°', 'unit': 'fahrenheit'}} execution_results = [get_weather_api_result] def add_execution_results(messages: List[Dict[str, Any]], execution_results: List[Dict[str, Any]]): content = "\n".join([f"\n{json.dumps(result)}" for result in execution_results]) messages.append({"role": "user", "content": content}) return messages messages = add_execution_results(messages, execution_results) inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device) outputs = model.generate( inputs, max_new_tokens=512, do_sample=False, num_return_sequences=1, eos_token_id=tokenizer.eos_token_id, ) response = tokenizer.decode(outputs[0][len(inputs[0]) :], skip_special_tokens=True) print(response) ```` Then you should be able to see the following output: ``` The current temperature in Seattle is 62 degrees in Fahrenheit. ``` # License Katanemo Arch-Function collection is distributed under the [Katanemo license](https://huggingface.co/katanemolabs/Arch-Function-3B/blob/main/LICENSE).