EVAL.md

BFCL-Hi Evaluation

Overview

BFCL-Hi (Berkeley Function-Calling Leaderboard - Hindi) is a Hindi adaptation of the BFCL benchmark, designed to evaluate the function-calling capabilities of Large Language Models in Hindi. This benchmark assesses models' ability to understand function descriptions and generate appropriate function calls based on natural language instructions.

Evaluation Workflow

BFCL-Hi follows the BFCL v2 evaluation methodology from the original GitHub repository, utilizing the same framework for assessing function-calling capabilities.

Evaluation Steps

1. Load the dataset (see important note below about dataset loading)
2. Generate model responses with function calls based on the prompts
3. Evaluate function-calling accuracy using the BFCL v2 evaluation scripts
4. Obtain metrics including execution accuracy, structural correctness, and other BFCL metrics

Important: Dataset Loading

⚠️ DO NOT use the HuggingFace load_dataset method to load the BFCL-Hi dataset.

The dataset files are hosted on HuggingFace but are not compatible with the HuggingFace datasets package. This is consistent with the English version of the dataset.

Recommended Approach:

• Download the JSON files directly from the HuggingFace repository
• Load them manually using standard JSON loading methods
• Follow the BFCL v2 repository's data loading methodology

Implementation

Please follow the same methodology as BFCL v2 (English) as documented in the official resources below.

Setup and Usage

Step 1: Installation

Clone the Gorilla repository and install dependencies:

git clone https://github.com/ShishirPatil/gorilla.git
cd gorilla/berkeley-function-call-leaderboard
pip install -r requirements.txt

Step 2: Prepare Your Dataset

• Place your dataset files in the appropriate directory
• Follow the data format specifications from the English BFCL v2

Step 3: Generate Model Responses

Run inference to generate function calls from your model:

python openfunctions_evaluation.py \
  --model <model_name> \
  --test-category <category> \
  --num-gpus <num_gpus>

Key Parameters:

• --model: Your model name or path
• --test-category: Category to test (e.g., all, simple, multiple, parallel, etc.)
• --num-gpus: Number of GPUs to use

For Hindi (BFCL-Hi):

• Ensure you load the Hindi version of the dataset
• Modify the inference code according to your model and hosted inference framework

Available Test Categories in BFCL-Hi:

• simple: Single function calls
• multiple: Multiple function calls
• parallel: Parallel function calls
• parallel_multiple: Combination of parallel and multiple function calls
• relevance: Testing function relevance detection
• irrelevance: Testing irrelevant function call handling

Step 4: Evaluate Results

Evaluate the generated function calls against ground truth:

python eval_runner.py \
  --model <model_name> \
  --test-category <category>

This will:

• Parse the generated function calls
• Compare with ground truth
• Calculate accuracy metrics
• Generate detailed error analysis

Step 5: View Results

Results will be saved in the output directory with metrics including:

• Execution Accuracy: Whether the function call executes correctly
• Structural Correctness: Whether the function call structure is valid
• Argument Accuracy: Whether arguments are correctly formatted
• Overall Score: Aggregated performance metric

You can also create custome Evaluation Script based on the above for more control over the evaluation process.

Official BFCL v2 Resources

• GitHub Repository: Berkeley Function-Calling Leaderboard

  • Complete evaluation framework and scripts
  • Dataset loading instructions
  • Evaluation metrics implementation

• BFCL v2 Documentation: BFCL v2 Release

  • Overview of v2 improvements and methodology

• Gorilla Project: https://gorilla.cs.berkeley.edu/

  • Main project page with additional resources

• Research Paper: Gorilla: Large Language Model Connected with Massive APIs

  • Patil et al., arXiv:2305.15334 (2023)