Changelog

All notable changes to the Berkeley Function Calling Leaderboard will be documented in this file.

• [Jan 20, 2025] #887: Add new model speakleash/Bielik-11B-v2.3-Instruct to the leaderboard.
• [Jan 18, 2025] #888: Add the following new models to the leaderboard:
  • NovaSky-AI/Sky-T1-32B-Preview
  • Qwen/QwQ-32B-Preview
• [Jan 12, 2025] #881: Fix Nova handler for consecutive user prompt issue.
• [Jan 11, 2025] : Add new model ZJared/Haha-7B to the leaderboard.
• [Jan 4, 2025] #865: Fix a copy-paste issue in live_parallel_multiple_9-8-0 that caused a misalignment between the question and the possible answer.
• [Jan 3, 2025] #864: Add support for pre-existing completion endpoints, allowing users to skip the local vLLM/SGLang server setup (using the --skip-server-setup flag) and point the generation pipeline to an existing OpenAI-compatible endpoint via VLLM_ENDPOINT and VLLM_PORT.
• [Jan 3, 2025] #859: Rename directories: proprietary_model -> api_inference, oss_model -> local_inference for better clarity.
• [Dec 29, 2024] #857: Add new model DeepSeek-V3 to the leaderboard.
• [Dec 29, 2024] #855: Add new model mistralai/Ministral-8B-Instruct-2410 to the leaderboard.
• [Dec 22, 2024] #838: Fix parameter type mismatch error in possible answers.
  • Simple: 2 affected
  • Multiple: 1 affected
  • Parallel: 6 affected
  • Parallel Multiple: 4 affected
  • Live Simple: 4 affected
  • Live Multiple: 26 affected
  • Live Parallel: 2 affected
• [Dec 22, 2024] #843: Add the following new models to the leaderboard:
  • gemini-2.0-flash-exp-FC
  • gemini-2.0-flash-exp
  • gemini-exp-1206-FC
  • gemini-exp-1206
• [Dec 21, 2024] #849: Use N/A in score report for unevaluated categories to distinguish from categories where the model actually scored a 0
• [Dec 21, 2024] #848: Improves behavior for generation and evaluation pipeline. When executable categories are involved and API keys are not provided in the .env file, instead of throwing an error, the affected categories will now be skipped. This enhancement provides a smoother experience for first-time users.
• [Dec 21, 2024] #847: Add new model watt-ai/watt-tool-8B and watt-ai/watt-tool-70B to the leaderboard.
• [Dec 20, 2024] #842: Add the following new models to the leaderboard:
  • Qwen/Qwen2.5-0.5B-Instruct
  • Qwen/Qwen2.5-3B-Instruct
  • Qwen/Qwen2.5-14B-Instruct
  • Qwen/Qwen2.5-32B-Instruct
• [Dec 18, 2024] #840: Add the following new models to the leaderboard:
  • o1-2024-12-17-FC
  • o1-2024-12-17
• [Dec 16, 2024] #837: Add the following new models to the leaderboard:
  • meta-llama/Llama-3.3-70B-Instruct-FC
  • meta-llama/Llama-3.3-70B-Instruct
• [Dec 13, 2024] #832: Add the following new models to the leaderboard:
  • MadeAgents/Hammer2.1-7b
  • MadeAgents/Hammer2.1-3b
  • MadeAgents/Hammer2.1-1.5b
  • MadeAgents/Hammer2.1-0.5b
• [Dec 11, 2024] #826, #829: Fix enum type mismatch error in function doc for live categories.
  • Live Simple: 7 affected
  • Live Multiple: 176 affected
  • Live Parallel Multiple: 3 affected
  • Live Irrelevance: 70 affected
• [Dec 9, 2024] #822: Add the following new models to the leaderboard:
  • gpt-4o-2024-11-20
  • gpt-4o-2024-11-20-FC
• [Dec 4, 2024] #815: Add the following new models to the leaderboard:
  • nova-pro-v1.0
  • nova-lite-v1.0
  • nova-micro-v1.0
• [Dec 3, 2024] #810: Add new model grok-beta to the leaderboard.
• [Dec 2, 2024] #809: Resolve issue in Gemini model when no model output.
• [Dec 2, 2024] #808: Improve latency measurement accuracy.
• [Nov 26, 2024] #755: Add new model palmyra-x-004 to the leaderboard.
• [Nov 25, 2024] #718: Add new model openbmb/MiniCPM3-4B-FC to the leaderboard.
• [Nov 25, 2024] #697: Add the following new models to the leaderboard:
  • deepseek-ai/DeepSeek-V2.5
  • deepseek-ai/DeepSeek-Coder-V2-Instruct-0724
  • deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
  • deepseek-ai/DeepSeek-V2-Chat-0628
  • deepseek-ai/DeepSeek-V2-Lite-Chat
• [Nov 25, 2024] #787: Add new model Qwen/Qwen2.5-72B-Instruct to the leaderboard.
• [Nov 24, 2024] #743: Add support for regeneration, specific test entry IDs, and custom directory locations:
  • Introduce the --allow-overwrite flag for the generate command to enable regeneration of test entries even if they already exist.
  • Add a new --run-ids flag for the generate command, allowing execution of specific test entry IDs from test_case_ids_to_generate.json.
    • Note: This cannot be used together with --test-category.
    • Test ids needs to be the exact same as the ones in the dataset. Example: "simple": ["simple_10", "simple_53"].
  • Add --score-dir and --result-dir options for generate and evaluate commands, enabling custom paths for result and score directories relative to the project root berkeley-function-call-leaderboard directory.
• [Nov 22, 2024] #777, #778, #881: Fix dataset entries where the function doc contains illegal Python parameter names (such as class). 55 entries are affected.
• [Nov 19, 2024] #750: Add the following new models to the leaderboard:
  • claude-3-5-haiku-20241022
  • claude-3-5-haiku-20241022-FC
  • claude-3-5-sonnet-20241022
  • claude-3-5-sonnet-20241022-FC
• [Nov 18, 2024] #736: Add the option to additionally log the evaluation results to WandB artifacts. User can enable this feature by providing the entity and project name in WANDB_BFCL_PROJECT in the .env file.
• [Nov 18, 2024] #768, #770: Resolve issues in Gemini models (FC mode) related to handling scenarios with no tools available and cases where the model output is empty.
• [Nov 17, 2024] #767: Fix price and latency calculation. A merge conflict results in a duplicate line, and counting the input and output token for each entry multiple times.
• [Nov 15, 2024] #762: Supply data_multi_turn.csv for multi-turn evaluation results
• [Nov 14, 2024] #760, #761: Upstream google-cloud-aiplatform library fixed typecasting bugs in Function Calling. Updated to version 1.72.0 and remove the workaround patch introduced in #648.
• [Nov 14, 2024] #747: Minor Grammatical Corrections to DEFAULT_SYSTEM_PROMPT that is supplied to all prompting models.
• [Nov 13, 2024] #737, #739, #740, #763, #772, #789, #804: Bug fix in the dataset and possible answers for the live and multi-turn categories.
• [Nov 11, 2024] #746: Improve inference log readability; inference log is now included as part of the model result file. For details on how to interpret the inference log, please refer to the LOG_GUIDE.md.
• [Nov 9, 2024] #749: Remove Llama-3.2-3B-Instruct-FC and Llama-3.2-1B-Instruct-FC from the leaderboard. According to the official Llama documentation, these models perform function calling using the prompt-style chat template rather than the specialized function-calling format.
• [Nov 8, 2024] #720: Add new model BitAgent/GoGoAgent to the leaderboard.
• [Oct 30, 2024] #725, #733: Update evaluation metric for multi-turn categories:
  • Introduce a new response-based checker, which works alongside with the existing state-based checker.
    • The new checker compares the model’s execution result against the ground truth execution result, ensuring that the model’s result encompasses the ground truth (i.e., ground truth must be a strict subset of the model result).
    • It complements the state-based checker, which doesn't work well when the functions don't directly alter the state. For example, it's unclear whether the model actually invoked get_zipcode_by_city or estimate_distance by just using the state-based checker.
    • Any multi turn entry will now only be marked correct if it passes both the state and response checkers.
  • Remove the irrelevance detection for multi-turn categories.
    • Instead of checking if the model produces no output in a turn with missing function/parameter information, we now assess whether the model can perform correctly once the missing information is provided.
  • A few dataset entries have been modified to align with these changes.
• [Oct 30, 2024] #719, #722, #723, #728, #732: Bug fix in the dataset and ground truth for the multi-turn categories.
• [Oct 17, 2024] #683: Bug fix for the multi turn categories for ambiguity in action intention and function parameters.
• [Oct 17, 2024] #709: Rephrase question prompt for Java and JavaScript categories to improve clarity and action intent.
• [Oct 17, 2024] #708: Update the ground truth for the REST category to be up-to-date with the latest API response structure.
• [Oct 16, 2024] #701: Bug fix the multi turn function source code for TravelAPI.
• [Oct 16, 2024] #696: Add the following new models to the leaderboard:
  • google/gemma-2-2b-it
  • google/gemma-2-9b-it
  • google/gemma-2-27b-it
• [Oct 16, 2024] #661: Bug fix in the dataset and possible answers.
  • Irrelevance: 1 affected
  • Parallel Multiple: 2 affected
  • Live Simple: 104 affected
  • Live Multiple: 547 affected
  • Live Parallel: 11 affected
  • Live Parallel Multiple: 17 affected
• [Oct 11, 2024] #667: Add the following new models to the leaderboard:
  • MadeAgents/Hammer2.0-7b
  • MadeAgents/Hammer2.0-3b
  • MadeAgents/Hammer2.0-1.5b
  • MadeAgents/Hammer2.0-0.5b
• [Oct 10, 2024] #621, #675: Add a basic command-line interface for ease of use.
• [Oct 5, 2024] #633: Add new model openbmb/MiniCPM3-4B to the leaderboard.
• [Oct 5, 2024] #642: Add the following new models to the leaderboard:
  • Qwen/Qwen2.5-7B-Instruct
  • Qwen/Qwen2.5-1.5B-Instruct
  • Qwen/Qwen2-7B-Instruct
  • Qwen/Qwen2-1.5B-Instruct
• [Oct 4, 2024] #653: Add new model Team-ACE/ToolACE-8B to the leaderboard.
• [Oct 4, 2024] #671: Speed up locally-hosted model's inference process by parallelizing the inference requests.
• [Sept 27, 2024] #640: Add the following new models to the leaderboard:
  • microsoft/Phi-3.5-mini-instruct
  • microsoft/Phi-3-medium-128k-instruct
  • microsoft/Phi-3-medium-4k-instruct
  • microsoft/Phi-3-small-128k-instruct
  • microsoft/Phi-3-small-8k-instruct
  • microsoft/Phi-3-mini-128k-instruct
  • microsoft/Phi-3-mini-4k-instruct
• [Sept 25, 2024] #660: Bug fix in parse_nested_value function to handle nested dictionary values properly.
• [Sept 24, 2024] #657: Add the following new models to the leaderboard:
  • meta-llama/Llama-3.2-1B-Instruct
  • meta-llama/Llama-3.2-1B-Instruct-FC
  • meta-llama/Llama-3.2-3B-Instruct
  • meta-llama/Llama-3.2-3B-Instruct-FC
  • meta-llama/Llama-3.1-8B-Instruct
  • meta-llama/Llama-3.1-8B-Instruct-FC
  • meta-llama/Llama-3.1-70B-Instruct
  • meta-llama/Llama-3.1-70B-Instruct-FC
• [Sept 24, 2024] #648: Add the following new models to the leaderboard:
  • gemini-1.5-pro-002
  • gemini-1.5-pro-002-FC
  • gemini-1.5-pro-001
  • gemini-1.5-pro-001-FC
  • gemini-1.5-flash-002
  • gemini-1.5-flash-002-FC
  • gemini-1.5-flash-001
  • gemini-1.5-flash-001-FC
  • gemini-1.0-pro-002
  • gemini-1.0-pro-002-FC
• [Sept 19, 2024] #644: BFCL V3 release:
  • Introduce new multi-turn dataset and state-based evaluation metric
  • Separate ast_checker and executable_checker for readability
  • Several outdated or deprecated models will be excluded from the leaderboard and replaced with their updated successors to improve the leaderboard's overall maintainability.
  • Switch to use vllm serve for OSS model inference
• [Sept 13, 2024] #638: Fix prompt formatting issue for THUDM/glm-4-9b-chat.
• [Sept 12, 2024] #635: Add new models o1-preview-2024-09-12 and o1-mini-2024-09-12 to the leaderboard.
• [Sept 8, 2024] #627 Add new model MadeAgents/Hammer-7b to the leaderboard.
• [Sept 7, 2024] #626: Fix prompt format for Llama models.
• [Sept 4, 2024] #623: Fix decoding issue in the NvidiaHandler; remove duplicate ArcticHandler class.
• [August 29, 2024] #616: Add the following new models to the leaderboard:
  • Salesforce/xLAM-7b-r
  • Salesforce/xLAM-8x7b-r
  • Salesforce/xLAM-8x22b-r
• [August 28, 2024] #565, #612: Packagerize the BFCL pipeline for easier deployment and maintenance.
• [August 27, 2024] #608: Bug fix in the dataset and possible answers.
  • simple: 16 affected
  • multiple: 5 affected
• [August 23, 2024] #600: Bug fix in the dataset and possible answers.
  • simple: 12 affected
  • multiple: 3 affected
  • parallel: 3 affected
  • parallel multiple: 6 affected
• [August 22, 2024] #593:
  • Move formatting instructions and function documentation to system prompt instead of user prompt in the message section. All prompting models are affected.
  • Bug fix in the dataset and possible answers.
    • irrelevance: 1 affected
    • live_irrelevance: 1 affected
    • live_simple: 1 affected
    • live_parallel: 3 affected
• [August 19, 2024] #580: Introduce BFCL V2 Live dataset, featuring user-contributed live prompts and function docs. To read more about the composition and construction of this dataset, please refer to our blog. All CLI commands have been updated to support the new dataset.
• [August 8, 2024] #574: Set temperature to 0.001 for all models for consistency and reproducibility.
• [August 7, 2024] #571: Support parallel inference for hosted models. User can specify the number of threads to use for parallel inference by setting the --num-threads flag. The default is 1, which means no parallel inference.
• [August 6, 2024] #569, #570, #573: Add the following new models to the leaderboard:
  • open-mistral-nemo-2407
  • open-mistral-nemo-2407-FC
  • open-mixtral-8x22b
  • open-mixtral-8x22b-FC
  • open-mixtral-8x7b
  • gpt-4o-mini-2024-07-18
  • gpt-4o-mini-2024-07-18-FC
  • gpt-4o-2024-08-06
  • gpt-4o-2024-08-06-FC
  • meetkai/functionary-medium-v3.1-FC
  • meetkai/functionary-small-v3.1-FC
  • meetkai/functionary-small-v3.2-FC
• [August 5, 2024] #568: Rephrase the question prompt for the executable_parallel_function category to remove potentially misleading information implying multi-turn function calls.
• [August 4, 2024] #557: Bug fix in the possible answers.
  • simple: 7 affected
  • multiple function: 3 affected
  • parallel function: 5 affected
  • parallel multiple function: 6 affected
  • executable parallel function: 1 affected
  • javascript: 3 affected
• [July 26, 2024] #549: Fix js_type_converter.py to properly handle JavaScript array value inside dictionary.
• [July 25, 2024] #532, #543, #556, #542: Add the following new models to the leaderboard:
  • Salesforce/xLAM-7b-fc-r
  • Salesforce/xLAM-1b-fc-r
  • yi-large-fc
  • NousResearch/Hermes-2-Pro-Llama-3-8B
  • NousResearch/Hermes-2-Pro-Llama-3-70B
  • NousResearch/Hermes-2-Theta-Llama-3-8B
  • NousResearch/Hermes-2-Theta-Llama-3-70B
• [July 22, 2024] #540: Chore: Improve handling of vLLM's cleanup phase error by combining all selected test categories into one single task to submit to the vLLM server.
• [July 21, 2024] #538, #545: Fix language_specific_pre_processing and convert_to_tool function to properly handle pre-processing for prompts and function docs in Java and JavaScript test categories. All entries in these categories are affected.
• [July 20, 2024] #537: Update generation script for locally-hosted OSS model to use single-node multi-GPU inference method (tensor parallel). Ray is not used anymore.
• [July 16, 2024] #525, #536: Add new model ibm-granite/granite-20b-functioncalling to the leaderboard.
• [July 10, 2024] #522: Bug fix in the evaluation dataset for Executable Parallel Multiple category. This includes updates to both prompts and function docs. 2 entries are affected.
• [July 8, 2024] #516: Fix double-casting issue in model_handler for Java and JavaScript test categories.
• [July 7, 2024] #504, #505, #506, #508, #512, #517: Make BFCL user-friendly and easy to extend.
• [July 6, 2024] #423 and #503: Bug fix in possible answers for the AST evaluation dataset (parallel category: 14 affected; parallel_multiple category: 25 affected).
• [July 5, 2024] #496: Updates to API status checks. Checking the health of executable APIs is now off by default. Further, even when triggered, un-healthy APIs will not terminate the evaluation process. Users can enable this feature by setting the --api-sanity-check flag or -c for short. The previous --skip-api-sanity-check or -s flag is now deprecated.
• [July 3, 2024] #489: Add new model nvidia/nemotron-4-340b-instruct to the leaderboard.
• [July 2, 2024] #474: Add new model THUDM/glm-4-9b-chat to the leaderboard.
• [June 18, 2024] #470: Add new model firefunction-v2-FC to the leaderboard.
• [June 15, 2024] #437: Fix prompting issues for Nexusflow-Raven-v2 (FC).
• [June 7, 2024] #407, #462: Update the AST evaluation logic to allow the use of int values for Python parameters expecting float values. This is to accommodate the Python auto-conversion feature from int to float.
• [May 14, 2024] #426:
  • Add the following new models to the leaderboard:
    • gpt-4o-2024-05-13
    • gpt-4o-2024-05-13-FC
    • gemini-1.5-pro-preview-0514
    • gemini-1.5-flash-preview-0514
  • Update price for the following models:
    • All Gemini Series
    • Claude-2.1 (Prompt) and Claude-instant-1.2 (Prompt)
    • Mistral-large and Mistral-Small
    • GPT-3.5-Turbo-0125
• [May 8, 2024] #406 and #421: Update the gemini_handler.py to better handle parallel function calls for Gemini models.
• [May 6, 2024] #412: Bug fix in evaluation dataset for AST categories. This includes updates to both prompts and function docs.
• [May 2, 2024] #405: Bug fix in the possible answers for the AST Simple evaluation dataset. Prompt and function docs are not affected.
• [April 28, 2024] #397: Add new model snowflake/arctic to the leaderboard. Note that there are multiple ways to inference the model, and we choose to do it via Nvidia API catalog.
• [April 27, 2024] #390: Bug fix in cost and latency calculation for open-source models, which are now all calculated when serving the model with vLLM using 8 V100 GPUs for consistency. $$\text{Cost} = \text{Latency per 1000 function call} * (\text{8xV100 azure-pay-as-you-go-price per hour / 3600})$$
• [April 25, 2024] #386: Add 5 new models to the leaderboard: meta-llama/Meta-Llama-3-8B-Instruct, meta-llama/Meta-Llama-3-70B-Instruct, gemini-1.5-pro-preview-0409, command-r-plus, command-r-plus-FC.
• [April 19, 2024] #377:
  • Bug fix for the evaluation dataset in the executable test categories. This includes updates to both prompts and function docs.
  • The evaluation_result field has been removed to accommodate the variability in API execution results across different evaluation runs. Instead, a human-verified ground_truth is now included for the executable test categories. During each evaluation run, evaluation_result is generated anew using the ground_truth, and then compared against the model output.
  • A stricter metric has been adopted when using the structural_match (aka. type match) evaluation criteria ---- For list results, the lengths are compared; for dict results, the keys are matched. This is to account for the fast-changing nature of some of the real-time API results while ensuring the evaluation remains meaningful.
  • Added another evaluation criteria real_time_match for the executable category, which is a looser form of exact_match specifically for numerical execution results. The execution result must be within a certain percentage threshold (20%) from the expected result to accommodate the live updates of API responses. User can change this threshold value in eval_checker_constant.py.
• [April 18, 2024] #375: A more comprehensive API sanity check is included; the APIs that are invoked during the non-REST executable evaluation process will also be checked for their availability before running the evaluation. Also, add support for the shortcut -s for the --skip-api-sanity-check flag, based on the community feedback.
• [April 16, 2024] #366: Switch to use Anthropic's new Tool Use Beta tools-2024-04-04 when generating Claude 3 FC series data. gpt-4-turbo-2024-04-09 and gpt-4-turbo-2024-04-09-FC are also added to the leaderboard.
• [April 11, 2024] #347: Add the 95th percentile latency to the leaderboard statistics. This metric is useful for understanding the latency distribution of the models, especially the worst-case scenario.
• [April 10, 2024] #339: Introduce REST API sanity check for the REST executable test category. It ensures that all the API endpoints involved during the execution evaluation process are working properly. If any of them are not behaving as expected, the evaluation process will be stopped by default as the result will be inaccurate. Users can choose to bypass this check by setting the --skip-api-sanity-check flag or -s for short.
• [April 9, 2024] #338: Bug fix in the evaluation datasets (including both prompts and function docs). Bug fix for possible answers as well.
• [April 8, 2024] #330: Fixed an oversight that was introduced in #299. For function-calling (FC) models that cannot take float type in input, when the parameter type is a float, the evaluation procedure will convert that type to number in the model input and mention in the parameter description that This is a float type value.. An additional field format: float will also be included in the model input to make it clear about the type. Updated the model handler for Claude, Mistral, and OSS to better parse the model output.
• [April 8, 2024] #327: Add new model NousResearch/Hermes-2-Pro-Mistral-7B to the leaderboard.
• [April 3, 2024] #309: Bug fix for evaluation dataset possible answers. Implement string standardization for the AST evaluation pipeline, i.e. removing white spaces and a subset of punctuations (,./-_*^) to make the AST evaluation more robust and accurate. Fixed AST evaluation issue for type tuple. Add 2 new models meetkai/functionary-small-v2.4 (FC), meetkai/functionary-medium-v2.4 (FC) to the leaderboard.
• [April 1, 2024] #299: Leaderboard update with new models (Claude-3-Haiku, Databrick-DBRX-Instruct), more advanced AST evaluation procedure, and updated evaluation datasets. Cost and latency statistics during evaluation are also measured. We also released the manual that our evaluation procedure is based on, available here.
• [Mar 11, 2024] #254: Leaderboard update with 3 new models: Claude-3-Opus-20240229 (Prompt), Claude-3-Sonnet-20240229 (Prompt), and meetkai/functionary-medium-v2.2 (FC)
• [Mar 5, 2024] #237 and 238: leaderboard update resulting from #223; 3 new models: mistral-large-2402, gemini-1.0-pro, and google/gemma-7b-it.
• [Feb 29, 2024] #223: modifications to REST evaluation.