🦍 Gorilla: Large Language Model Connected with Massive APIs

Berkeley Function-Calling Leaderboard (BFCL) aims to provide a thorough study of the function-calling capability of different LLMs. It consists of 2k question-function-answer pairs with multiple languages (python, java, javascript, restAPI), diverse application domains and complex use cases (multiple function calls where the LLM needs to select one or more functions from multiple functions provided, and parallel function calls that the LLM needs to make multiple function calls together). We also investigate function relevance detection, to determine how the model will react when the provided function is not suitable to answer the user's question (in such case an "Error Message will be provided"). In more detail, BFCL includes 100 Java, 50 JavaScript, 70 REST API, 100 SQL, and 1,680 Python on various simple, parallel, multiple, executable functions calling scenarios as well as function relevance detection.

The leaderboard is shown below in the Figure, we can see that the latest checkpoint of GPT-4 (from OpenAI) leads the evaluation, with the open-source model (OpenFunctions-v2), Mistral-medium model (from Mistral AI) and Claude-2.1 (from Anthropic) following close behind. This blog post includes more information on the dataset, the evaluation methodology, some common failure patterns, and more!

LLMs' performance on Berkeley Function-Calling Leaderboard (BFCL)

To improve our understanding and visualization of the outcomes, we have introduced an interactive wagon wheel tool that allows users to compare various models. This comparison is organized into nine distinct categories: function relevance detection, AST (Abstract Syntax Tree) tree analysis, and execution function call verification across simple, multiple, and parallel multiple function scenarios. Through this approach, it becomes evident that tests reveal unsatisfactory performance by the models. Specifically, in simple function calling, both proprietary and open-source models exhibit comparable performance. However, when it comes to handling multiple and parallel function calls, the GPT-series models demonstrate superior performance over their open-source counterparts.

Detailed analysis using Berkeley Function-Calling Leaderboard (BFCL) Wagon Chart

The Gorilla OpenFunctions evaluation dataset grows from its previous OpenFunctions-v0's 100 data points to 2,000 data points! Beyond improvements in quality, the expanded dataset demonstrates diversity in:

• Domains of functions documentation
• Number of function documents and function call(s) pairs
• Data types of different programming languages
• Executability of real-world examples

Our evaluation JSON functions are scraped and generated from different sources of websites. We intentionally include domains like using functions related to Mathematics-Algebra, Sports-Soccer, Finance-Mortgage, etc. We include 40 sub-domains of functions within our generic evaluations. This allows us to understand the model performance not just in data-abundant domains like computing, and cloud, but also in niche domains like sports, and law.

Berkeley Function-Calling Leaderboard (BFCL) Data Composition

We break down the majority of the evaluation into two categories:

• Python: Simple Function, Multiple Function, Parallel Function, Parallel Multiple Function
• Non-Python: Chatting Capability, Function Relevance Detection, REST API, SQL, Java, Javascript

Simple Function: Single function evaluation contains the simplest but most commonly seen format, where the user supplies a single JSON function document, with one and only one function call will be invoked.

Multiple Function: Multiple function category contains a user question that only invokes one function call out of 2 to 4 JSON function documentations. The model needs to be capable of selecting the best function to invoke according to user-provided context.

Parallel Function: Parallel function is defined as invoking multiple function calls in parallel with one user query. The model needs to digest how many function calls need to be made and the question to model can be a single sentence or multiple sentence.

Parallel Multiple Function: Parallel Multiple function is the combination of parallel function and multiple function. In other words, the model is provided with multiple function documentation, and each of the corresponding function calls will be invoked zero or more times.

Each category has both AST and its corresponding executable evaluations. In the executable evaluation data, we manually write Python functions drawing inspiration from free REST API endpoints (e.g. get weather) and functions (e.g. linear regression) that compute directly. The executable category is designed to understand whether the function call generation is able to be stably utilized in applications utilizing function calls in the real world.

While the previous categories consist of the majority of our evaluations, we include other specific categories, namely Chatting Capability, Function Relevance Detection, REST API, SQL, Java, and JavaScript, to evaluate model performance on diverse scenarios and support of multiple programming languages, and are resilient to irrelevant questions and function documentations.

Chatting Capability: In Chatting Capability, we design scenarios where no functions are passed in, and the users ask generic questions - this is similar to using the model as a general-purpose chatbot. We evaluate if the model is able to output chat messages and recognize that it does not need to invoke any functions. Note the difference with “Relevance” where the model is expected to also evaluate if any of the function inputs are relevant or not. We include this category for internal model evaluation and exclude the statistics from the live leaderboard. We currently are working on a better evaluation of chat ability and ensuring the chat is relevant and coherent with users' requests and open to suggestions and feedback from the community.

Function Relevance Detection: In function relevance detection, we design scenarios where none of the provided functions are relevant and supposed to be invoked. We expect the model's output to be no function call. This scenario provides insight into whether a model will hallucinate on its function and parameter to generate function code despite lacking the function information or instructions from the users to do so.

REST API: A majority of the real-world API calls are from REST API calls. Python mainly makes REST API calls through requests.get(), requests.post(), requests.delete(), etc that are included in the Python requests library. GET requests are the most common ones used in the real world. As a result, we include real-world GET requests to test the model's capabilities to generate executable REST API calls through complex function documentations, using requests.get() along with the API's hardcoded URL and description of the purpose of the function and its parameters. Our evaluation includes two variations. The first type requires passing the parameters inside the URL, called path parameters, for example, the {Year} and {CountryCode} in GET /api/v3/PublicHolidays/{Year}/{CountryCode}. The second type requires the model to put parameters as key/value pairs into the params and/or headers of requests.get(.). For example, params={'lang': 'fr'} in the function call. The model is not given which type of REST API call it is going to make but needs to make a decision on how it's going to be invoked.

For REST API, we use an executable evaluation to check for the executable outputs' effective execution, response type and response JSON key consistencies. On the AST, we chose not to perform AST evaluation on REST mainly because of the immense number of possible answer for complicated defined APIs that enumeration of all possible answers is exhaustive.

SQL: SQL evaluation data includes our customized sql.execute functions that contain sql_keyword, table_name, columns, and conditions. Those four parameters provide the necessary information to construct a simple SQL query like SELECT column_A from table_B where column_C == D Through this, we want to see if through function calling, SQL query can be reliably constructed and utilized rather than training a SQL-specific model. In our evaluation dataset, we restricted the scenarios and supported simple keywords, including SELECT, INSERT INTO, UPDATE, DELETE, and CREATE. We included 100 examples for SQL AST evaluation. Note that SQL AST evaluation will not be shown in our leaderboard calculations. We use SQL evaluation to test the generalization ability of function calling for programming languages that are not included in the training set for Gorilla OpenFunctions-v2. We opted to exclude SQL performance from the AST evaluation in the BFCL due to the multiplicity of methods to construct SQL function calls achieving identical outcomes. We're currently working on a better evaluation of SQL and are open to suggestions and feedback from the community. Therefore, SQL has been omitted from the current leaderboard to pave the way for a more comprehensive evaluation in subsequent iterations.

Java + Javascript: Despite function calling formats being the same across most programming languages, each programming language has language-specific types. For example, Java has the HashMap type. The goal of this test category is to understand how well the function calling model can be extended to not just Python type but all the language-specific typings. We included 100 examples for Java AST evaluation and 70 examples for Javascript AST evaluation.

The categories outlined above provide insight into the performance of different models across popular API call scenarios, offering valuable perspectives on the potential of function-calling models.

We've performed a hierarchical categorization on our existing categories to have nine categories showcased in our Berkeley Function-Calling Leaderboard BFCL, which we group by on both evaluation method (AST or execution) and type of functions (simple, parallel, multiple, parallel multiple functions). Here, we display a table organizing counts of evaluation data points of each leaderboard category, which is composed of more granular categories listed in the blog. Specifically, we categorize REST executable evaluation as Simple Function (Evaluation by Executing APIs) because we considered cases where one REST API call is being called. For Java + Javascript evaluation, we categorize these into Simple Function (Abstract Syntax Tree (AST) Evaluation) since our current version of the evaluation set didn't include multiple, parallel, and parallel multiple cases of diverse programming languages.

The final counts of each of the nine categories shown in BFCL with the composition of more granular types are shown in the following table

We use two popular methods to evaluate the accuracy of the model-generated answers: AST evaluation and Executable evaluation. Ideally one should use execution evaluation, but when we evaluate the answers, not all the results are easily executable (e.g., Java functions). So we use the AST as a complement to the execution evaluation.

• Abstract Syntax Tree (AST) Evaluation
• Executable Function Evaluation

For simple function evaluations, the evaluation process focuses on comparing a single model output function against its function doc and possible answers. Here is a flow chart that shows the step-by-step evaluation process.

Evaluating function calls. The function description, and possible answer are used to evaluate the model's output function.

Parsing Function Through AST

The evaluation process starts by parsing the function call using the AST tree.

Example:

calculate_triangle_area(base=10, height=5)

Parse:

Module(body=[
        Expr(value=List(elts=[
            Call(
                func=Name(id='calculate_triangle_area', ctx=Load()), 
                args=[], 
                keywords=[
                    keyword(arg='base', value=Constant(value=10)), 
                    keyword(arg='height', value=Constant(value=5))
                ]
            )
        ], ctx=Load()))
    ], type_ignores=[])

Function Matching

The procedure first extracts the function name and verifies that it is consistent with the one in possible answer.

• Here, note that the function name may contain a .. Given that certain models (e.g. OpenAI series) may not support the dots . in their input, we substitute dots with underscores _ in function names when inferencing the models to generate function outputs. This substitution is repeated during the result evaluation phase.

Required Parameters Matching

Then, it extracts the arguments from the AST and check if each parameter can be found and exact matched in possible answers.

• The evaluation process ensures all required parameters, as identified by the "required" attribute in the function documentation, are present in the model output.
• It also ensures that only parameters exist in the function doc are used, flagging the model hallucination outputs.

Parameter Type & Value Matching

The evaluation process is strict on typing. Here are the acceptable answers for each data type:

• For bool:
  • The procedure checks the direct matching of boolean values, and doesn't allow leniency on the string versions of boolean values.
• For integer, float:
  • If the function documentation specifies that it's a float, then it should be a float in the model output (such as 5.0); an int (such as 5) will not be correct.
• For List, Tuple:
  • Order matters and the elements must match exactly. For example, [1,2,3] is not equal to [2,3,1]. So for questions where the order in the list doesn't matter, permutations of the possible answer are used to accommodate this situation.
  • Note that the type match extends recursively for nested data structures (List or Tuple), where both the outer type and the inner types of elements must match the specified requirements.
• For String:
  • The evaluation process is case-insensitive.
  • All strings will be standardized before checking. This applies to both the model output and the possible answers.
  • All white space is removed.
  • A subset of punctuations ,./-_*^ are removed to make the evaluation more robust and accurate.
  • Possible date ["20th June", "2023-06-20", "06/20/2023", "Jun.20, 2023"]
  • Possible Location ["New York City", "NYC"]
  • Possible Anything ["Manchester United", "Man United", "Man U", "MUFC"]
• For Dict:
  • The evaluation focuses on the key presence and the accuracy of associated values as per the possible answers.
  • Ordering within dictionaries is not considered due to they are inherently unordered.
• For lists of dictionaries:
  • While the ordering of dictionaries is considered (since it's a List), the order of key-value pairs within each dictionary is not.
• Handling Optional Parameters in Function Calls:
  • For parameters that are truly optional (in other words, the function doc didn't list that parameter as required and the possible answer contains the empty string ""), the model can choose to use the default value for that parameter, or not provide value. Both are considered correct.
  • For parameters that aren't listed as required in the function doc, and the possible answer didn't contain the empty string "", then it is not truly optional. The prompt must imply that we should use some value different from the default for that parameter. So in this case, the model needs to explicitly provide the correct value; using default value or not providing value would be marked as wrong.
• When the parameter value is a variable mentioned in the prompt, the procedure special handles. For example, if the type of parameter param1 is type integer, the user could provide param1=value1 (where value1 is a concrete value of type integer), or param1=variable1 (where variable1 is a variable that holds an integer value). So the evaluation takes into account both scenarios when checking the type.
• For the Java and JavaScript test categories, all the parameter values (in the input) are in the string format. The evaluation procedure will call the java_type_converter or the js_type_converter helper method to convert the string version of a Java/JS type into their corresponding Python format, since we perform checking with Python data types. For example, a HashMap in Java will be converted to a dictionary in Python. During this process, the converter will also perform the type checking for those parameters; e.g. if the parameter should be a long type in Java, the converter will check to make sure that the string input does have an "L" at the end (because otherwise, it wouldn't be a valid Java long).

Here are some examples of possible answers:

{"calculate_triangle_area": {"base": [10], "height": [5], "unit": ["units", "unit"]}}

{"predict_house_price": {"bedrooms": [3], "bathrooms": [2], "area": [1800], "location": ["San Francisco", "San Francisco, CA"]}}

Cost & Latency

In our recent update, we have also paid close attention to the cost and latency.

• For models from service providers such as OpenAI, Mistral, Google, Anthropic, and etc:
  • Latency: we measure the latency by timing each request to the endpoint ignoring the function document preprocessing time.
  • Cost: we follow the formula to derive the cost per 1000 function callings.
• For models that we evaluate using local hosting. This includes Deepseek, Gemma, and etc.:
  • Latency: For ~7B Models, we evaluated on single A100 40GB. For ~40B Models, we evaluated on 4*A100 40GB. Since we batched and evaluated the model, we derive latency by dividing the total time by the number of evaluation dataset entries.
  • Cost: Since the open source model does not have a price tag, we estimate the cost by:

For firefunctions-v1, Nexusflow-Raven-v2, and Meetkai-Functionary, we use their endpoints while their services are free, so we did not include the cost for their models.

This specific piece of insight will help individuals or enterprises to decide which model to incorporate based on the demand and budget.

When to function-call (tool-call) and when to prompt? 🖊️

1. GPTs' function documents are difficult to format and their typings are restrictive in real-world scenarios.

   "Function": 
   {
       "name": "calculate_binomial_probability",
       ...
       "parameters": 
       {
           "type": "object", 
           "properties": 
           {
               "number_of_trials": 
               {
                   "type": "integer", 
                   "description": "The total number of trials."
               },
               "number_of_successes": 
               {
                   "type": "integer", 
                   "description": "The desired number of successful outcomes."
               },
               "probability_of_success": 
               {
                   "type": "float", 
                   The probability of a successful outcome on any given trial.", 
                   "default": 0.5
               }
               ...
           }
           "required": ["number_of_trials", "number_of_successes"]
       }
   }

   In this case, we need to manually convert float into number to make the function OpenAI compatible. In addition to that, numbers convey less information compared to floats in terms of precision and type consistency.

   In Gorilla OpenFunctions-v2, we improve the flexibility of the function documents by not restricting the typing of the parameters. In other words, the user can supply Tuple, Float, and even language-specific types like Hashmap and Linked List in Java!

2. GPT underperforms in scenarios where the parameters are not immediately available in the user question but instead require some implicit conversions. Here is an example:

   "Function": 
   {
       "name": "finance.predict_future_value",
       ...
       "parameters": 
       {
           "type": "object", 
           "properties": 
           {
               "present_value": 
               {
                   "type": "number", 
                   "description": "The present value of the investment."
               },
               
               "annual_interest_rate": 
               {
                   "type": "number", 
                   "description": "The annual interest rate of the investment."
               },
               
               "compounding_periods_per_year": 
               {
                   "type": "integer", 
                   "description": "The number of times that interest is compounded per year.", 
               },
               "time_years": 
               {
                   "type": "integer", 
                   "description": "The investment horizon in years."
               }
               ...
           }
           "required": ["present_value", "annual_interest_rate", "time_years"]
       }
   }

   Questions: Predict the future value of a $5000 investment with an annual interest rate of 5% in 3 years with monthly compounding.

   GPT-4 output:
   [{
       "name": "finance.predict_future_value",
       "parameters": 
       {
           "present_value": 5000,
           "annual_interest_rate": 5,
           "compounding_periods_per_year": 12,
           "time_years": 3
       }
   }]

   Gorilla-openfunctions-v2 output:
   [{
       "name": "finance.predict_future_value",
       "parameters": 
       {
           "present_value": 5000,
           "annual_interest_rate": 0.05,
           "compounding_periods_per_year": 12,
           "time_years": 3
       }
   }]

3. Chat Models tend to generate malformed function calls in which parameters can be extracted but not executable

   Example: mistral-medium generates results like solve\\_quadratic\\_equation(a=2, b=6, c=5). With gorilla-openfunctions-v2, we are able to directly output solve_quadratic_equation(a=3, b=2, c=1) which is executable upon receiving the result.

4. REST API missing URLs:

   A discrepancy arises due to the absence of the required URL in REST API requests made by GPT-4 for fetching weather data. While the GPT-4 output omits the necessary URL, the Gorilla Openfunctions-v2 model successfully includes the correct API endpoint, enabling it to successfully execute and retrieve the requested weather information for the specified coordinates and forecast period.

   "User": "Can you fetch me the weather data for the coordinates 
   37.8651 N, 119.5383 W, including the hourly forecast for temperature, 
   wind speed, and precipitation for the next 10 days?"
   
   "Function": 
   {
       "name": "requests.get",
       ...
       "parameters": 
       {
           "type": "object", 
           "properties": 
           {
               "url": 
               {
                   "type": "string", 
                   "description": "The API endpoint for fetching
                   weather data from the Open-Meteo API for the 
                   given latitude and longitude, default 
                   https://api.open-meteo.com/v1/forecast"
               }
               ...
           }
       }
   }

   GPT-4 output:
   {
       "name": "requests.get",
       "parameters": {
           "url": "Missing",
           "params": 
           {
               "latitude": "37.8651",
               "longitude": "-119.5383",
               "forecast_days": 10
           },
       }
   }

   Gorilla-Openfunctions-v2 output:
   {
       "name": "requests.get",
       "parameters": {
           "url": "https://api.open-meteo.com/v1/forecast",
           "params": 
           {
               "latitude": "37.8651",
               "longitude": "-119.5383",
               "forecast_days": 10
           },
       }
   }

We provide a comprehensive and systematic evaluation of LLMs for function calling with Gorilla Open Functions Leaderboard. The studies here suggest that in terms of simple function calling (without complex planning and chained function calling), finetuning an open-source can be as effective as propriety models. Furthermore, we provide Gorilla Open Functions v2, an open-source model that can help users with building AI applications with function calling and interacting with JSON compatible output.

We hope you enjoyed this blog post. We would love to hear from you on Discord, Twitter (#GorillaLLM), and GitHub.

If you would like to cite Gorilla:
@inproceedings{berkeley-function-calling-leaderboard,
  title={Berkeley Function Calling Leaderboard},
  author={Fanjia Yan and Huanzhi Mao and Charlie Cheng-Jie Ji and Tianjun Zhang and Shishir G. Patil and Ion Stoica and Joseph E. Gonzalez},
  year={2024},
  howpublished={\url{https://gorilla.cs.berkeley.edu/blogs/8_berkeley_function_calling_leaderboard.html}},
}