README.md

metadata

title: Les Audits d'Affaires - Leaderboard
emoji: ⚖️
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.34.2
app_file: app.py
pinned: false
license: mit
short_description: Leaderboard français pour LLMs sur droit des affaires

Les Audits d'Affaires - Leaderboard

Performance dashboard for LLMs on French business law benchmark with HuggingFace Datasets integration.

🚀 Setup Complete!

HuggingFace Datasets

• Requests: legmlai/laal-requests - tracks evaluation requests
• Results: legmlai/laal-results - stores evaluation results

Current Models (with 0 scores)

1. Qwen/Qwen3-14B (Alibaba)
2. jpacifico/Chocolatine-2-14B-Instruct-v2.0.3 (jpacifico)
3. meta-llama/Llama-3.1-8B-Instruct (Meta)

🏃‍♂️ Quick Start

Prerequisites

export HF_TOKEN=your_huggingface_token

Run Leaderboard

cd les-audites-affaires-leadboard
source venv/bin/activate
python app.py

The leaderboard will be available at: http://127.0.0.1:7860

📁 Project Structure

Core Files

• app.py - Main leaderboard application with HuggingFace integration
• dataset_manager.py - HuggingFace datasets management
• requirements.txt - Python dependencies

Setup Scripts

• create_datasets.py - Initialize HuggingFace datasets
• setup_initial_models.py - Add initial models with 0 scores

✨ Features

Live Leaderboard

• Real-time data from HuggingFace datasets
• Automatic ranking and scoring
• Category-wise performance breakdown
• Interactive comparison charts

Model Submissions

• Submit models directly through the UI
• Automatic request tracking
• Email notifications for updates

Pipeline Status Tracking

• 📊 Real-time status: Pending, Processing, Completed, Failed
• 📋 Recent evaluation requests table
• 🔄 Pipeline progress monitoring
• ⏳ Request status updates with emojis

Data Management

• All data stored in HuggingFace datasets
• Refresh button for live updates
• Persistent across sessions
• Status update automation ready

Token Management

• Graceful handling when HF_TOKEN not configured
• Clear instructions for Hugging Face Spaces deployment
• Read-only mode when token unavailable

🔄 Next Steps

1. Deploy to Hugging Face Spaces: Upload to HF Spaces with HF_TOKEN secret
2. Connect Evaluation Harness: Use automation_example.py as integration guide
3. Real Evaluations: Replace 0 scores with actual evaluation results
4. Webhooks: Set up automatic updates from evaluation pipeline
5. Enhanced Analytics: Add more detailed performance breakdowns

📊 Dataset Schema

Requests Dataset

• request_id, model_name, model_provider, request_type
• request_status, contact_email, request_timestamp

Results Dataset

• result_id, request_id, model_name, model_provider
• overall_score, score_action_requise, score_delai_legal
• score_documents_obligatoires, score_impact_financier
• score_consequences_non_conformite, evaluation_timestamp, is_published

---

Created by: Mohamad Alhajar (legml.ai)
License: MIT