2008robocode-crypto/code-generation-system

1"""2Evaluation test dataset: 10 real prompts + 10 edge cases3Used to measure success rate, retries, and reliability metrics.4"""5 6TEST_PROMPTS = {7    "real_products": [8        {9            "id": "real_1",10            "name": "CRM System",11            "prompt": "Build a CRM with login, contacts, dashboard, role-based access, and premium plan with payments. Admins can see analytics."12        },13        {14            "id": "real_2",15            "name": "E-commerce Platform",16            "prompt": "Create an e-commerce platform with product listing, shopping cart, checkout, payment processing, order tracking, and admin inventory management."17        },18        {19            "id": "real_3",20            "name": "Project Management Tool",21            "prompt": "Build a project management tool with tasks, teams, real-time collaboration, comments, file attachments, and Gantt chart visualization."22        },23        {24            "id": "real_4",25            "name": "Social Network",26            "prompt": "Create a social network with user profiles, friend connections, feed, messaging, notifications, and content sharing."27        },28        {29            "id": "real_5",30            "name": "Booking System",31            "prompt": "Build a booking platform for services with service listings, availability calendar, booking management, payments, and customer reviews."32        },33        {34            "id": "real_6",35            "name": "Learning Platform",36            "prompt": "Create an online learning platform with courses, lessons, quizzes, student progress tracking, certificates, and instructor dashboards."37        },38        {39            "id": "real_7",40            "name": "Chat Application",41            "prompt": "Build a messaging app with direct messages, group chats, file sharing, user presence, and message history."42        },43        {44            "id": "real_8",45            "name": "Analytics Dashboard",46            "prompt": "Create a business analytics dashboard with data visualization, custom reports, filters, export functionality, and real-time metrics."47        },48        {49            "id": "real_9",50            "name": "Healthcare Portal",51            "prompt": "Build a healthcare patient portal with appointment scheduling, medical records, doctor messaging, prescription management, and billing."52        },53        {54            "id": "real_10",55            "name": "HR Management System",56            "prompt": "Create an HR management system with employee records, leave management, attendance tracking, payroll integration, and performance reviews."57        }58    ],59    "edge_cases": [60        {61            "id": "edge_1",62            "category": "vague",63            "name": "Super Vague",64            "prompt": "Build something useful"65        },66        {67            "id": "edge_2",68            "category": "vague",69            "name": "Unclear Requirements",70            "prompt": "App with features and things and stuff"71        },72        {73            "id": "edge_3",74            "category": "conflicting",75            "name": "Conflicting Requirements",76            "prompt": "Build a lightweight app with heavy real-time analytics and complex machine learning. Keep it simple but feature-rich with enterprise-grade security and minimal code."77        },78        {79            "id": "edge_4",80            "category": "conflicting",81            "name": "Contradictory Roles",82            "prompt": "Create a system where guests can manage admin settings but admins have read-only access."83        },84        {85            "id": "edge_5",86            "category": "incomplete",87            "name": "Missing Core Details",88            "prompt": "Build an app"89        },90        {91            "id": "edge_6",92            "category": "incomplete",93            "name": "No Business Rules",94            "prompt": "System with users, products, and orders"95        },96        {97            "id": "edge_7",98            "category": "ambiguous",99            "name": "Ambiguous Scope",100            "prompt": "Build social platform with all standard features"101        },102        {103            "id": "edge_8",104            "category": "ambiguous",105            "name": "Undefined Entities",106            "prompt": "Create a marketplace for things between people with verification and payments"107        },108        {109            "id": "edge_9",110            "category": "complex",111            "name": "Over-specified",112            "prompt": "Build a system with microservices architecture, GraphQL API, React frontend, PostgreSQL database, Redis caching, Kafka message queue, Docker containerization, Kubernetes orchestration, CI/CD pipeline, and monitoring with Prometheus and Grafana. Also include blockchain integration for immutable audit logs."113        },114        {115            "id": "edge_10",116            "category": "technical",117            "name": "Technical Jargon",118            "prompt": "Implement CRUD operations for normalization with ACID compliance, horizontal scalability, and auto-sharding across multiple zones"119        }120    ]121}122 123 124def get_test_dataset():125    """Get all test prompts."""126    return TEST_PROMPTS127 128 129def get_real_prompts():130    """Get just the real product prompts."""131    return TEST_PROMPTS["real_products"]132 133 134def get_edge_cases():135    """Get just the edge case prompts."""136    return TEST_PROMPTS["edge_cases"]137 138 139def get_prompt_by_id(prompt_id: str):140    """Get a specific test prompt."""141    for category in TEST_PROMPTS.values():142        for item in category:143            if item["id"] == prompt_id:144                return item145    return None146