Enhances Berkeley Function Calling Leaderboard with parameter counts and efficiency metrics for identifying the most cost-effective AI models.

BFCL-Model-Efficiency-Analyzer.user.js

// ==UserScript==
// @name         BFCL Parameter Analysis
// @namespace    http://tampermonkey.net/
// @version      2025-05-13
// @description  Extract parameter counts and add efficiency metrics to Berkeley Function Calling Leaderboard
// @author       Fredrik Angelsen
// @match        https://gorilla.cs.berkeley.edu/leaderboard.html
// @icon         https://www.google.com/s2/favicons?sz=64&domain=berkeley.edu
// @grant        none
// ==/UserScript==

(function() {
    'use strict';

    // Wait for the table to be fully loaded
    const checkForTable = setInterval(() => {
        const table = document.getElementById('leaderboard-table');
        if (table && table.querySelector('tbody tr')) {
            clearInterval(checkForTable);
            enhanceLeaderboard();
        }
    }, 1000);

    function enhanceLeaderboard() {
        // Get the table header and add new column headers
        const headerRow = document.querySelector('#leaderboard-head tr:last-child');

        // Add new header cells
        const paramHeader = document.createElement('th');
        paramHeader.className = 'column-header';
        paramHeader.textContent = 'Params (B)';
        paramHeader.id = 'param-col';
        paramHeader.addEventListener('click', () => sortTableByColumn('param'));
        paramHeader.innerHTML += '<span></span>'; // For sorting indicator

        const efficiencyHeader = document.createElement('th');
        efficiencyHeader.className = 'column-header';
        efficiencyHeader.textContent = 'Acc/Param';
        efficiencyHeader.id = 'efficiency-col';
        efficiencyHeader.addEventListener('click', () => sortTableByColumn('efficiency'));
        efficiencyHeader.innerHTML += '<span></span>'; // For sorting indicator

        headerRow.appendChild(paramHeader);
        headerRow.appendChild(efficiencyHeader);

        // Process each row in the table
        const rows = document.querySelectorAll('#leaderboard-table tbody tr');
        rows.forEach(row => {
            // Get model name and accuracy
            const modelCell = row.querySelector('td:nth-child(3)');
            const modelName = modelCell.textContent;
            const accuracy = parseFloat(row.querySelector('td:nth-child(2)').textContent);

            // Extract parameter count
            const params = extractParameterCount(modelName);

            // Calculate efficiency (accuracy per billion parameters)
            const efficiency = params > 0 ? (accuracy / params).toFixed(2) : 'N/A';

            // Create and append parameter cell
            const paramCell = document.createElement('td');
            paramCell.textContent = params > 0 ? params : 'Unknown';
            paramCell.dataset.value = params;
            row.appendChild(paramCell);

            // Create and append efficiency cell
            const efficiencyCell = document.createElement('td');
            efficiencyCell.textContent = efficiency;
            efficiencyCell.dataset.value = efficiency !== 'N/A' ? efficiency : -1;
            row.appendChild(efficiencyCell);
        });

        // Add explanation of the new metrics
        const explanation = document.createElement('p');
        explanation.innerHTML = '<b>Params (B)</b> = Parameter count in billions. <b>Acc/Param</b> = Accuracy per billion parameters (higher is better).';
        const leaderboardDiv = document.getElementById('leaderboard');
        leaderboardDiv.querySelector('.col-md-12').appendChild(explanation);
    }

    function extractParameterCount(modelName) {
        // Extract parameter count using regex
        const patterns = [
            /(\d+(?:\.\d+)?)b/i, // matches 32b, 3.5b, etc.
            /(\d+(?:\.\d+)?)B/, // matches 70B, 3.5B, etc.
            /-(\d+(?:\.\d+)?)b-/i, // matches -32b-, -3.5b-, etc.
            /-(\d+(?:\.\d+)?)B-/, // matches -70B-, -3.5B-, etc.
            /(\d+(?:\.\d+)?)b-/i, // matches 32b-, 3.5b-, etc.
            /(\d+(?:\.\d+)?)B-/, // matches 70B-, 3.5B-, etc.
            /-(\d+(?:\.\d+)?)b/i, // matches -32b, -3.5b, etc.
            /-(\d+(?:\.\d+)?)B/, // matches -70B, -3.5B, etc.
        ];

        // Try each regex pattern
        for (const pattern of patterns) {
            const match = modelName.match(pattern);
            if (match && match[1]) {
                return parseFloat(match[1]);
            }
        }

        return 0; // Unknown size
    }

    function sortTableByColumn(column) {
        const table = document.getElementById('leaderboard-table');
        const tbody = table.querySelector('tbody');
        const rows = Array.from(tbody.querySelectorAll('tr'));

        // Get the current sort direction or set default
        const header = document.getElementById(`${column}-col`);
        const sortSpan = header.querySelector('span');
        const isAscending = sortSpan.textContent === ' 🔼';

        // Update sorting indicators
        document.querySelectorAll('#leaderboard-head th span').forEach(span => {
            span.textContent = '';
        });

        // Set the appropriate sort direction indicator
        sortSpan.textContent = isAscending ? ' 🔽' : ' 🔼';

        rows.sort((a, b) => {
            let aValue, bValue;

            if (column === 'param') {
                aValue = parseFloat(a.querySelector('td:nth-last-child(2)').dataset.value) || 0;
                bValue = parseFloat(b.querySelector('td:nth-last-child(2)').dataset.value) || 0;
            } else if (column === 'efficiency') {
                aValue = parseFloat(a.querySelector('td:last-child').dataset.value) || -1;
                bValue = parseFloat(b.querySelector('td:last-child').dataset.value) || -1;
            }

            return isAscending ? aValue - bValue : bValue - aValue;
        });

        // Re-add sorted rows to the table
        rows.forEach(row => tbody.appendChild(row));
    }
})();

To use this script:

1. Install the Tampermonkey browser extension
2. Click the Tampermonkey icon in your browser
3. Select "Create new script"
4. Delete any existing code and paste this entire script
5. Save (Ctrl+S or File > Save)
6. Visit https://gorilla.cs.berkeley.edu/leaderboard.html to see the enhanced table