Unlocking Gemma 3’s Potential: How This New AI Model Stacks Up in Real Tests

import os
import time
import json
import torch
import argparse
from datetime import datetime
from dotenv import load_dotenv
from transformers import AutoTokenizer, Gemma3ForCausalLM
from huggingface_hub import login
import gc

# Load environment variables from .env file
load_dotenv()

# Get Hugging Face token from environment variables
HF_TOKEN = os.getenv("HF_TOKEN")
if not HF_TOKEN:
    raise ValueError("HF_TOKEN not found in environment variables. Please check your .env file.")

# Login to Hugging Face Hub using the token
login(token=HF_TOKEN)
print("Successfully logged in to Hugging Face Hub")

# Test prompts - using the same as in your original script
TEST_PROMPTS = {
    "basic_qa": [
        {
            "name": "transfer_learning",
            "prompt": "Explain the concept of transfer learning in a concise paragraph."
        },
        {
            "name": "ada_lovelace",
            "prompt": "Who was Ada Lovelace, and why is she significant in computing history?"
        },
        {
            "name": "overfitting_definition",
            "prompt": "What does the term 'overfitting' mean in machine learning?"
        },
        {
            "name": "gradient_descent",
            "prompt": "Briefly describe gradient descent and how it works."
        },
        {
            "name": "supervised_vs_unsupervised",
            "prompt": "Name two major differences between supervised and unsupervised learning."
        }
    ],
    "code_generation": [
        {
            "name": "sum_even_numbers",
            "prompt": "Write a Python function that takes a list of integers and returns the sum of all even numbers."
        },
        {
            "name": "debug_snippet",
            "prompt": (
                "Here is a Python snippet that's causing an error. Please fix it:\n\n"
                "```python\ndef greet(name):\n print(Hello, name)\n```"
            )
        },
        {
            "name": "factorial_function",
            "prompt": "Create a Python function that calculates the factorial of a given integer."
        },
        {
            "name": "cpp_hello_world",
            "prompt": "Generate a minimal C++ code snippet that prints 'Hello World' to the console."
        },
        {
            "name": "optimize_loops",
            "prompt": (
                "Optimize this Python code for performance:\n\n"
                "```python\n"
                "for i in range(10000):\n"
                " for j in range(10000):\n"
                " pass\n"
                "```"
            )
        }
    ],
    "summarization": [
        {
            "name": "internet_history",
            "prompt": (
                "Summarize the following text in under 50 words:\n\n"
                "The Internet's origins trace back to the 1960s, during the Cold War, when the U.S. Department of Defense's "
                "Advanced Research Projects Agency (ARPA) funded ARPANET, a project designed to create a decentralized "
                "communication network that could withstand a nuclear attack. In the 1970s, TCP/IP protocols were developed, "
                "laying the foundation for the modern Internet. The 1980s saw the rise of the Domain Name System (DNS) and "
                "the National Science Foundation Network (NSFNET), which expanded access beyond military and academic "
                "institutions. The World Wide Web, invented by Tim Berners-Lee in 1989, revolutionized the Internet by "
                "introducing hypertext and a user-friendly interface. The 1990s marked the commercialization of the Internet, "
                "with the emergence of web browsers and e-commerce. Today, the Internet is a global network connecting billions "
                "of devices and shaping nearly every aspect of modern life."
            )
        },
        {
            "name": "climate_change_impact",
            "prompt": (
                "Summarize the following text in under 50 words:\n\n"
                "Climate change, driven by the increasing concentration of greenhouse gases in the atmosphere, is causing "
                "significant and widespread impacts on the planet. Rising global temperatures lead to melting polar ice caps "
                "and glaciers, resulting in sea-level rise and coastal flooding. Extreme weather events, such as hurricanes, "
                "droughts, and wildfires, are becoming more frequent and intense. Changes in precipitation patterns disrupt "
                "agriculture and water resources, threatening food security. Ocean acidification, caused by increased "
                "absorption of carbon dioxide, harms marine ecosystems. These impacts have profound consequences for human "
                "societies, economies, and natural environments, necessitating urgent action to mitigate and adapt to climate "
                "change."
            )
        },
        {
            "name": "ai_healthcare",
            "prompt": (
                "Summarize the following text in under 50 words:\n\n"
                "Artificial intelligence (AI) is transforming healthcare by enabling faster and more accurate diagnoses, "
                "personalized treatments, and improved patient outcomes. AI-powered algorithms can analyze medical images, "
                "such as X-rays and MRIs, to detect diseases like cancer and Alzheimer's earlier than human experts. Machine "
                "learning models can predict patient risk for various conditions, allowing for proactive interventions. AI-driven "
                "drug discovery accelerates the development of new therapies. Virtual assistants and chatbots provide patients "
                "with 24/7 access to medical information and support. While AI offers immense potential, challenges remain, "
                "including data privacy concerns, regulatory hurdles, and the need for seamless integration into existing "
                "healthcare systems."
            )
        },
        {
            "name": "mobile_evolution",
            "prompt": (
                "Summarize the following text in under 50 words:\n\n"
                "The evolution of mobile technology has dramatically reshaped communication and information access. From bulky "
                "analog phones in the 1980s, mobile devices have transformed into powerful handheld computers. The introduction "
                "of 2G networks enabled text messaging and basic data services, while 3G brought faster internet speeds and "
                "multimedia capabilities. The advent of 4G LTE revolutionized mobile broadband, supporting high-definition "
                "video streaming and real-time applications. 5G technology promises even faster speeds, lower latency, and "
                "greater network capacity, enabling new applications like augmented reality and autonomous vehicles. The "
                "proliferation of smartphones and mobile apps has created a ubiquitous computing environment, impacting "
                "everything from social interactions to business operations."
            )
        },
        {
            "name": "renewable_energy",
            "prompt": (
                "Summarize the following text in under 50 words:\n\n"
                "The rise of renewable energy sources is a critical component of the global effort to combat climate change. "
                "Solar, wind, hydro, and geothermal power are increasingly replacing fossil fuels, reducing greenhouse gas "
                "emissions and air pollution. Technological advancements have driven down the cost of renewable energy, "
                "making it economically competitive with traditional sources. Government policies, such as subsidies and "
                "carbon pricing, are accelerating the transition to clean energy. Energy storage solutions, like batteries, "
                "are improving the reliability and grid integration of renewables. While challenges remain, including "
                "intermittency and infrastructure development, the momentum behind renewable energy is undeniable, paving the "
                "way for a sustainable energy future."
            )
        }
    ],
    "multilingual": [
        {
            "name": "spanish_question",
            "prompt": "¿Qué es la inteligencia artificial y por qué es importante?"
        },
        {
            "name": "french_question",
            "prompt": "Quels sont les avantages de l'apprentissage profond dans la vision par ordinateur?"
        },
        {
            "name": "chinese_question",
            "prompt": "在机器翻译中最大的挑战是什么？"
        },
        {
            "name": "arabic_question",
            "prompt": "ما هي تطبيقات الذكاء الاصطناعي في مجال الرعاية الصحية؟"
        },
        {
            "name": "german_question",
            "prompt": "Wie unterscheidet sich überwachtes Lernen vom unüberwachten Lernen?"
        }
    ]
}

def log_gpu_memory():
    """Log and return GPU memory usage in MB."""
    if torch.cuda.is_available():
        devices = list(range(torch.cuda.device_count()))
        memory_allocated = []
        
        for device in devices:
            mem_allocated = torch.cuda.memory_allocated(device) / (1024 ** 2)
            mem_reserved = torch.cuda.memory_reserved(device) / (1024 ** 2)
            memory_allocated.append(mem_allocated)
            print(f"GPU {device}: Allocated: {mem_allocated:.2f} MB, Reserved: {mem_reserved:.2f} MB")
        
        return sum(memory_allocated)
    else:
        print("No GPU available")
        return 0

def clear_gpu_memory():
    """Clear GPU memory cache."""
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        print("GPU memory cache cleared")
        gc.collect()
        return log_gpu_memory()
    else:
        print("No GPU available")
        return 0

def format_message(prompt):
    """Format a text prompt as a message for the model."""
    return [
        {
            "role": "user",
            "content": [{"type": "text", "text": prompt}]
        }
    ]

def test_model(model_id, output_dir="./results"):
    """Test the 1B model using direct model loading instead of pipeline."""
    print(f"\n{'='*50}\nTESTING MODEL: {model_id}\n{'='*50}")
    
    # Create output directory
    os.makedirs(output_dir, exist_ok=True)
    
    # Create results file with timestamp
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    model_name = model_id.split("/")[-1]
    result_file = os.path.join(output_dir, f"{model_name}_{timestamp}.json")
    results = []
    
    try:
        # Log initial GPU state
        print("Initial GPU memory state:")
        initial_gpu_mem = log_gpu_memory()
        
        # Load model and tokenizer
        print(f"Loading model: {model_id}")
        start_time = time.time()
        
        print("Device set to use cuda:0")
        device = "cuda:0" if torch.cuda.is_available() else "cpu"
        
        # Load the model and tokenizer
        tokenizer = AutoTokenizer.from_pretrained(model_id, token=HF_TOKEN)
        model = Gemma3ForCausalLM.from_pretrained(
            model_id, 
            device_map="auto",
            torch_dtype=torch.bfloat16,
            token=HF_TOKEN
        )
        
        load_time = time.time() - start_time
        print(f"Model loaded in {load_time:.2f} seconds")
        
        # Log GPU memory after loading
        print("GPU memory after model load:")
        gpu_after_load = log_gpu_memory()
        
        # Run tests for each category and prompt
        for category, prompts in TEST_PROMPTS.items():
            print(f"\n--- Testing {category} prompts ---")
            
            for prompt_data in prompts:
                prompt_name = prompt_data["name"]
                prompt_text = prompt_data["prompt"]
                
                print(f"Running: {prompt_name}")
                
                try:
                    # Start timing
                    inference_start = time.time()
                    
                    # Format as a chat message for instruction-tuned models
                    messages = format_message(prompt_text)
                    
                    # Apply chat template to format for model
                    inputs = tokenizer.apply_chat_template(
                        messages,
                        add_generation_prompt=True,
                        tokenize=True,
                        return_dict=True,
                        return_tensors="pt"
                    ).to(model.device)
                    
                    # Run inference
                    with torch.inference_mode():
                        outputs = model.generate(
                            **inputs,
                            max_new_tokens=512,
                            do_sample=True,
                            temperature=0.7,
                            top_p=0.9
                        )
                    
                    # Get just the generated part
                    input_length = inputs["input_ids"].shape[-1]
                    generated_tokens = outputs[0][input_length:]
                    
                    # Decode to get the text
                    response = tokenizer.decode(generated_tokens, skip_special_tokens=True)
                    
                    inference_time = time.time() - inference_start
                    print(f"Completed in {inference_time:.2f} seconds")
                    
                except Exception as e:
                    print(f"Error during inference: {str(e)}")
                    import traceback
                    traceback.print_exc()
                    response = f"ERROR: {str(e)}"
                    inference_time = time.time() - inference_start
                
                # Record results
                result_entry = {
                    "model_id": model_id,
                    "category": category,
                    "test_name": prompt_name,
                    "prompt": prompt_text,
                    "load_time_s": load_time,
                    "inference_time_s": inference_time,
                    "gpu_after_load_mb": gpu_after_load,
                    "output": response
                }
                
                results.append(result_entry)
                
                # Save results after each prompt
                with open(result_file, 'w', encoding='utf-8') as f:
                    json.dump(results, f, indent=2, ensure_ascii=False)
        
        print(f"\nAll tests completed for {model_id}. Results saved to: {result_file}")
        
    except Exception as e:
        print(f"Error testing model {model_id}: {str(e)}")
        import traceback
        traceback.print_exc()
    
    finally:
        # Clean up
        if 'model' in locals():
            del model
        if 'tokenizer' in locals():
            del tokenizer
        gc.collect()
        clear_gpu_memory()
    
    return result_file

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Test Gemma 3 1B model")
    parser.add_argument(
        "--output_dir", 
        type=str, 
        default="./results",
        help="Directory to save results"
    )
    
    args = parser.parse_args()
    
    # Check CUDA availability
    if torch.cuda.is_available():
        print(f"CUDA available: {torch.cuda.is_available()}")
        print(f"CUDA version: {torch.version.cuda}")
        print(f"Number of GPUs: {torch.cuda.device_count()}")
        for i in range(torch.cuda.device_count()):
            print(f"GPU {i}: {torch.cuda.get_device_name(i)}")
    else:
        print("WARNING: CUDA not available. Tests will run on CPU.")
    
    # Test the 1B model
    test_model("google/gemma-3-1b-it", args.output_dir)

[
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "basic_qa",
    "test_name": "transfer_learning",
    "prompt": "Explain the concept of transfer learning in a concise paragraph.",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 10.08859658241272,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "Transfer learning is a powerful technique where you leverage knowledge gained from solving one problem and apply it to a different, but related, problem. Instead of training a model from scratch, you start with a pre-trained model – often trained on a massive dataset – and fine-tune it for your specific task. This saves time, resources, and often leads to better performance, especially when you have limited data for your target task. Essentially, it’s like using your existing driving skills to learn how to drive a truck!"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "basic_qa",
    "test_name": "ada_lovelace",
    "prompt": "Who was Ada Lovelace, and why is she significant in computing history?",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 35.09038472175598,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "Okay, let's dive into the fascinating life and significance of Ada Lovelace!\n\n**Who was Ada Lovelace?**\n\nAda Lovelace (1815-1852) was an English mathematician and writer, primarily known for her work on Charles Babbage's proposed Analytical Engine. She's often considered the **first computer programmer** – a truly groundbreaking and somewhat surprising designation for someone living in the 19th century.\n\n**Here's a breakdown of her life and contributions:**\n\n* **Family Background:** Ada was the only legitimate child of the famous Romantic poet Lord Byron and Anne Isabella Milbanke, a woman with a strong interest in mathematics and logic. Her parents separated when she was young, and her mother was determined to prevent her from following in her father's poetic footsteps. This fostered a very analytical and logical mind in Ada.\n* **Early Education & Interest in Numbers:** Ada received a rigorous education, including tutoring from prominent mathematicians like William Frend and Mary Somerville, a renowned science writer. She developed a keen interest in mathematics and logic from a young age.\n* **The Collaboration with Babbage:** In 1833, Ada met Charles Babbage, the inventor of the Difference Engine and, more importantly, the Analytical Engine – a mechanical general-purpose computer.  She became fascinated by the potential of the Analytical Engine.\n* **The Notes on the Engine:**  This is where her genius truly shines.  Ada translated an article by Italian military engineer Luigi Menabrea about the Analytical Engine.  But crucially, she didn't just translate; she *expanded* upon it. She added extensive notes, which were three times longer than the original article. These notes are what cemented her place in history.\n* **\"Notes\" – The Algorithm:**  Within her notes, Ada outlined an algorithm for the Analytical Engine to compute Bernoulli numbers. This is widely recognized as the **first algorithm intended to be processed by a machine**, essentially making her the first computer programmer. She understood the machine's potential to do *more* than just calculate numbers – it could manipulate symbols and create complex outputs.\n\n**Why is she Significant in Computing History?**\n\n* **Visionary Thinking:** Ada Lovelace saw beyond the basic calculation capabilities of Babbage’s machine. She recognized that the Analytical Engine could be used for a far broader range of tasks, including composing music, creating graphics, and potentially even simulating complex scientific processes. This was incredibly forward"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "basic_qa",
    "test_name": "overfitting_definition",
    "prompt": "What does the term 'overfitting' mean in machine learning?",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 35.07544755935669,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "Okay, let's break down the concept of \"overfitting\" in machine learning. It's a really important and often frustrating issue!\n\n**What is Overfitting?**\n\nIn simple terms, overfitting means that a machine learning model learns the **training data *too* well**. It essentially memorizes the specific examples in the training set instead of learning the underlying, general patterns that would allow it to accurately predict new, unseen data.\n\n**Here's a more detailed explanation:**\n\n1. **Training Data:** You feed a machine learning model a large dataset of examples – the \"training data.\" This is the data the model uses to learn.\n\n2. **Learning the Patterns:** The model analyzes this data and builds a model (a set of parameters or weights) that can accurately predict the outputs for the examples it's seen.\n\n3. **Overfitting Happens:** When a model overfits, it becomes incredibly specialized to the training data. It learns the noise, the random fluctuations, and the specific details of *that particular dataset*.  It performs exceptionally well on the data it was trained on, but it struggles to generalize to new, unseen data.\n\n**Think of it like this:**\n\n* **Good Learning:** Imagine you're teaching a student to identify different types of birds. You show them hundreds of pictures of robins, cardinals, and blue jays. The student learns the *specific features* of those birds – the color of their feathers, the shape of their beak, etc.  They'll be excellent at identifying robins, but they'll struggle to identify a new, slightly different type of bird.\n\n* **Overfitting:**  The model, instead of learning the *general characteristics* of birds (like their ability to fly, their diet, their habitat), just memorizes the exact details of the pictures it was trained on.  It might perform brilliantly on the training set, but its performance on new, unseen birds will be poor.\n\n\n**Key Indicators of Overfitting:**\n\n* **High Accuracy on Training Data:** The model performs exceptionally well on the data it was trained on.\n* **Low Accuracy on Test Data:** The model performs poorly on a separate set of data (the \"test data\") that it hasn't seen before.\n* **Complex Model:** Overfitted models often have a very large number of parameters (weights and biases).\n\n**How to Prevent Overfitting:**\n\n* **More Data:**  Often, the best solution is"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "basic_qa",
    "test_name": "gradient_descent",
    "prompt": "Briefly describe gradient descent and how it works.",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 33.1804141998291,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "Okay, here's a brief explanation of gradient descent:\n\n**What is Gradient Descent?**\n\nGradient descent is a **supervised learning algorithm** used to find the minimum of a function.  In the context of machine learning, it's used to train models (like neural networks) by adjusting their parameters (weights and biases) to minimize the error between the model's predictions and the actual values.\n\n**How it Works:**\n\n1. **Start Somewhere:** You begin with a random set of model parameters.\n2. **Calculate the Gradient:**  The \"gradient\" tells you the direction of the *steepest* increase in the error.  It points uphill.\n3. **Take a Step:** You move in the *opposite* direction of the gradient (downhill). This is done by adjusting the parameters slightly.\n4. **Repeat:** You repeat steps 2 and 3 until you reach a point where the error is minimized (or close to it).\n\n**Analogy:**\n\nImagine you're standing on a mountain and you want to get to the lowest point. You can't see the whole landscape, but you can feel the slope. Gradient descent is like taking small steps downhill, always moving in the direction where the slope is steepest.\n\n**Key Points:**\n\n* **Iterative:** Gradient descent is an iterative process – you keep repeating the steps until you find a good solution.\n* **Loss Function:** You use a \"loss function\" to measure how well your model is performing. The goal is to minimize this loss.\n\n---\n\n**Resources for Further Learning:**\n\n* **Towards Data Science - Gradient Descent:** [https://towardsdatascience.com/gradient-descent-introduction-and-how-it-works-684632a3897a](https://towardsdatascience.com/gradient-descent-introduction-and-how-it-works-684632a3897a)\n* **Google AI - Gradient Descent:** [https://ai.google/research/gradient-descent/](https://ai.google/research/gradient-descent/)\n\n\nWould you like me to elaborate on any specific aspect of gradient descent (e.g., different types of gradient descent, how it relates to optimization algorithms)?"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "basic_qa",
    "test_name": "supervised_vs_unsupervised",
    "prompt": "Name two major differences between supervised and unsupervised learning.",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 16.850188493728638,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "Okay, here are two major differences between supervised and unsupervised learning:\n\n1. **Presence of Labeled Data:**\n   * **Supervised Learning:** Requires *labeled* data. This means you have a dataset where each input data point is paired with the correct output or \"label\" that the algorithm should predict. Think of it like teaching a child – you show them examples with the correct answers.\n   * **Unsupervised Learning:** Doesn't require labeled data. Instead, it analyzes unlabeled data to discover patterns, structures, or groupings within the data itself.  It's like giving a child a box of blocks and letting them figure out how to organize them.\n\n2. **Goal/Objective:**\n   * **Supervised Learning:** The goal is to *predict* an output for new, unseen data based on the patterns it learned from the labeled training data. It’s about making accurate predictions.\n   * **Unsupervised Learning:** The goal is to *discover* hidden structures or relationships in the data. It's about exploring and understanding the data without a predefined target variable.\n\n\nDo you want me to elaborate on any of these differences, or perhaps provide examples of each?"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "code_generation",
    "test_name": "sum_even_numbers",
    "prompt": "Write a Python function that takes a list of integers and returns the sum of all even numbers.",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 35.06111741065979,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "```python\ndef sum_of_even_numbers(numbers):\n  \"\"\"\n  Calculates the sum of all even numbers in a list.\n\n  Args:\n    numbers: A list of integers.\n\n  Returns:\n    The sum of all even numbers in the list.  Returns 0 if the list is empty \n    or contains no even numbers.\n  \"\"\"\n  even_sum = 0\n  for number in numbers:\n    if number % 2 == 0:\n      even_sum += number\n  return even_sum\n\n# Example usage:\nnumbers1 = [1, 2, 3, 4, 5, 6]\nresult1 = sum_of_even_numbers(numbers1)\nprint(f\"Sum of even numbers in {numbers1}: {result1}\")  # Output: 12\n\nnumbers2 = [1, 3, 5, 7]\nresult2 = sum_of_even_numbers(numbers2)\nprint(f\"Sum of even numbers in {numbers2}: {result2}\")  # Output: 0\n\nnumbers3 = []\nresult3 = sum_of_even_numbers(numbers3)\nprint(f\"Sum of even numbers in {numbers3}: {result3}\")  # Output: 0\n\nnumbers4 = [2, 4, 6, 8]\nresult4 = sum_of_even_numbers(numbers4)\nprint(f\"Sum of even numbers in {numbers4}: {result4}\") # Output: 20\n```\n\nKey improvements and explanations:\n\n* **Clear Function Definition:**  The code defines a function `sum_of_even_numbers(numbers)` that takes a list of integers as input.\n* **Docstring:**  Includes a docstring that clearly explains what the function does, its arguments, and what it returns.  This is crucial for good code documentation.\n* **Initialization:** `even_sum = 0` initializes the sum to zero.  This is important; if the list is empty or contains no even numbers, the function will correctly return 0.\n* **Modulo Operator:**  `number % 2 == 0` efficiently checks if a number is even. The modulo operator (`%`) returns the remainder of a division. If the remainder when dividing by 2 is 0, the number"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "code_generation",
    "test_name": "debug_snippet",
    "prompt": "Here is a Python snippet that's causing an error. Please fix it:\n\n```python\ndef greet(name):\n print(Hello, name)\n```",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 6.167387008666992,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "```python\ndef greet(name):\n print(\"Hello, \" + name)\n```\n\n**Reasoning:**\n\nThe original code had a syntax error. The `print` statement was missing a space before the `name` variable.  Adding a space ensures that the string \"Hello, \" is concatenated correctly with the value of `name`.  The corrected code now correctly prints \"Hello, \" followed by the provided name.\n"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "code_generation",
    "test_name": "factorial_function",
    "prompt": "Create a Python function that calculates the factorial of a given integer.",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 34.81158089637756,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "```python\ndef factorial(n):\n  \"\"\"\n  Calculates the factorial of a non-negative integer.\n\n  Args:\n    n: A non-negative integer.\n\n  Returns:\n    The factorial of n, or 1 if n is 0.\n    Returns None if n is negative.\n  \"\"\"\n  if n < 0:\n    return None  # Factorial is not defined for negative numbers\n  elif n == 0:\n    return 1\n  else:\n    result = 1\n    for i in range(1, n + 1):\n      result *= i\n    return result\n\n# Example usage:\nnumber = 5\nfact = factorial(number)\n\nif fact is not None:\n  print(f\"The factorial of {number} is {fact}\")\nelse:\n  print(\"Factorial is not defined for negative numbers.\")\n\nnumber = 0\nfact = factorial(number)\nprint(f\"The factorial of {number} is {fact}\")\n\nnumber = -3\nfact = factorial(number)\nif fact is not None:\n  print(f\"The factorial of {number} is {fact}\")\nelse:\n  print(\"Factorial is not defined for negative numbers.\")\n```\n\nKey improvements and explanations:\n\n* **Handles Negative Input:** The function now explicitly checks if `n` is negative.  Factorial is not defined for negative numbers, so it returns `None` in this case. This is crucial for robustness.\n* **Handles Zero Input:**  The function correctly handles the case where `n` is 0, returning 1 (because 0! = 1).\n* **Clearer Logic:** The `else` block is more concise and readable.\n* **Iterative Approach:** Uses an iterative approach (a `for` loop) which is generally more efficient than recursion for factorial calculations in Python.  Recursive solutions can quickly lead to stack overflow errors for larger numbers.\n* **Docstring:**  Includes a comprehensive docstring explaining the function's purpose, arguments, and return value.  Good documentation is essential.\n* **Example Usage with Error Handling:** The example usage now demonstrates how to handle the `None` return value when the input is negative, providing a complete and understandable example.  It also includes a test case for 0.\n* **f-strings:** Uses f-strings for"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "code_generation",
    "test_name": "cpp_hello_world",
    "prompt": "Generate a minimal C++ code snippet that prints 'Hello World' to the console.",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 26.787078857421875,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "```cpp\n#include <iostream>\n\nint main() {\n  std::cout << \"Hello World\" << std::endl;\n  return 0;\n}\n```\n\n**Explanation:**\n\n* **`#include <iostream>`:** This line includes the `iostream` library, which provides input and output functionalities, specifically for using `std::cout`.\n* **`int main() { ... }`:** This is the main function where the program execution begins.\n* **`std::cout << \"Hello World\" << std::endl;`:**\n    * `std::cout`: This is the standard output stream object (typically the console).\n    * `<<`: This is the insertion operator, used to send data to the output stream.\n    * `\"Hello World\"`: This is the string literal that will be printed to the console.\n    * `std::endl`: This inserts a newline character into the output stream, moving the cursor to the next line.  It's equivalent to `\\n`.\n* **`return 0;`:** This line returns 0 from the `main` function, indicating that the program executed successfully.\n\n**How to compile and run this code:**\n\n1. **Save the code:** Save the code in a file named `hello.cpp` (or any name ending in `.cpp`).\n2. **Compile:** Open a terminal or command prompt and use a C++ compiler (like g++) to compile the code:\n   ```bash\n   g++ hello.cpp -o hello\n   ```\n   This command compiles `hello.cpp` and creates an executable file named `hello`.\n3. **Run:** Execute the compiled program:\n   ```bash\n   ./hello\n   ```\n   This will print \"Hello World\" to the console.\n"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "code_generation",
    "test_name": "optimize_loops",
    "prompt": "Optimize this Python code for performance:\n\n```python\nfor i in range(10000):\n for j in range(10000):\n pass\n```",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 35.97528553009033,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "The provided code is extremely inefficient.  It's a loop that iterates 10,000 times, doing nothing in between.  This is a classic example of a \"looping without purpose.\"\n\nHere's how to optimize it, along with explanations:\n\n**1. The Correct Approach:  Use a List Comprehension or a Generator Expression**\n\nThe most efficient way to achieve the same result (printing 10,000 numbers) is to use a list comprehension or a generator expression.  These are generally faster than explicit loops in Python.\n\n* **List Comprehension:**\n\n   ```python\n   numbers = [i * 10000 for i in range(10000)]\n   for num in numbers:\n       print(num)\n   ```\n\n* **Generator Expression:**\n\n   ```python\n   numbers = (i * 10000 for i in range(10000))\n   for num in numbers:\n       print(num)\n   ```\n\n   Generator expressions are memory-efficient because they produce values on demand, rather than creating a complete list in memory upfront.\n\n**2.  Why these are faster:**\n\n* **Python's Optimized Loops:** Python's `for` loops are highly optimized.  They are implemented in C and leverage low-level optimizations.\n* **List Comprehension/Generator Expressions:** These are often implemented in C, providing significant speed advantages over Python's interpreted loop execution.\n* **Avoiding Unnecessary Copies:**  The original code creates a new list in each iteration of the loop.  List comprehensions and generator expressions avoid this overhead.\n\n**3.  Alternative (More Complex, but potentially useful in certain situations):  Using NumPy (if you're working with numerical data)**\n\nIf you're working with numerical data and need to perform a numerical operation (like multiplication), NumPy is almost always the best choice.\n\n```python\nimport numpy as np\n\nnumbers = np.arange(10000) * 10000\nfor num in numbers:\n    print(num)\n```\n\n**4.  Why avoid the original loop (if you can't change the code):**\n\nThe original code is the *worst* possible way to accomplish this.  It's a waste of processing power.  If you *must* use the original loop, you could consider:\n\n"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "summarization",
    "test_name": "internet_history",
    "prompt": "Summarize the following text in under 50 words:\n\nThe Internet's origins trace back to the 1960s, during the Cold War, when the U.S. Department of Defense's Advanced Research Projects Agency (ARPA) funded ARPANET, a project designed to create a decentralized communication network that could withstand a nuclear attack. In the 1970s, TCP/IP protocols were developed, laying the foundation for the modern Internet. The 1980s saw the rise of the Domain Name System (DNS) and the National Science Foundation Network (NSFNET), which expanded access beyond military and academic institutions. The World Wide Web, invented by Tim Berners-Lee in 1989, revolutionized the Internet by introducing hypertext and a user-friendly interface. The 1990s marked the commercialization of the Internet, with the emergence of web browsers and e-commerce. Today, the Internet is a global network connecting billions of devices and shaping nearly every aspect of modern life.",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 5.49566912651062,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "Here’s a summary under 50 words:\n\nThe Internet began in the 1960s as ARPANET, spurred by the Cold War. TCP/IP, DNS, and the Web transformed it in the 70s and 80s, ultimately leading to its global expansion and commercialization today, connecting billions worldwide."
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "summarization",
    "test_name": "climate_change_impact",
    "prompt": "Summarize the following text in under 50 words:\n\nClimate change, driven by the increasing concentration of greenhouse gases in the atmosphere, is causing significant and widespread impacts on the planet. Rising global temperatures lead to melting polar ice caps and glaciers, resulting in sea-level rise and coastal flooding. Extreme weather events, such as hurricanes, droughts, and wildfires, are becoming more frequent and intense. Changes in precipitation patterns disrupt agriculture and water resources, threatening food security. Ocean acidification, caused by increased absorption of carbon dioxide, harms marine ecosystems. These impacts have profound consequences for human societies, economies, and natural environments, necessitating urgent action to mitigate and adapt to climate change.",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 4.616601467132568,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "Here’s a summary of the text in under 50 words:\n\nClimate change, fueled by greenhouse gases, is drastically impacting the planet. Rising temperatures cause sea-level rise, extreme weather, disrupted agriculture, and ocean acidification, threatening human societies and ecosystems. Urgent action is needed to mitigate and adapt to these challenges."
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "summarization",
    "test_name": "ai_healthcare",
    "prompt": "Summarize the following text in under 50 words:\n\nArtificial intelligence (AI) is transforming healthcare by enabling faster and more accurate diagnoses, personalized treatments, and improved patient outcomes. AI-powered algorithms can analyze medical images, such as X-rays and MRIs, to detect diseases like cancer and Alzheimer's earlier than human experts. Machine learning models can predict patient risk for various conditions, allowing for proactive interventions. AI-driven drug discovery accelerates the development of new therapies. Virtual assistants and chatbots provide patients with 24/7 access to medical information and support. While AI offers immense potential, challenges remain, including data privacy concerns, regulatory hurdles, and the need for seamless integration into existing healthcare systems.",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 3.955840587615967,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "Here’s a summary of the text in under 50 words:\n\nAI is revolutionizing healthcare with faster diagnoses, personalized treatments, and improved patient outcomes through image analysis, predictive modeling, and drug discovery. However, challenges like data privacy and integration remain crucial for successful implementation."
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "summarization",
    "test_name": "mobile_evolution",
    "prompt": "Summarize the following text in under 50 words:\n\nThe evolution of mobile technology has dramatically reshaped communication and information access. From bulky analog phones in the 1980s, mobile devices have transformed into powerful handheld computers. The introduction of 2G networks enabled text messaging and basic data services, while 3G brought faster internet speeds and multimedia capabilities. The advent of 4G LTE revolutionized mobile broadband, supporting high-definition video streaming and real-time applications. 5G technology promises even faster speeds, lower latency, and greater network capacity, enabling new applications like augmented reality and autonomous vehicles. The proliferation of smartphones and mobile apps has created a ubiquitous computing environment, impacting everything from social interactions to business operations.",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 4.659716367721558,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "Here’s a summary of the text in under 50 words:\n\nMobile technology has evolved dramatically, transitioning from bulky phones to powerful handheld devices. 2G, 3G, 4G LTE, and now 5G have fueled rapid advancements in connectivity, enabling new applications and transforming daily life through ubiquitous computing."
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "summarization",
    "test_name": "renewable_energy",
    "prompt": "Summarize the following text in under 50 words:\n\nThe rise of renewable energy sources is a critical component of the global effort to combat climate change. Solar, wind, hydro, and geothermal power are increasingly replacing fossil fuels, reducing greenhouse gas emissions and air pollution. Technological advancements have driven down the cost of renewable energy, making it economically competitive with traditional sources. Government policies, such as subsidies and carbon pricing, are accelerating the transition to clean energy. Energy storage solutions, like batteries, are improving the reliability and grid integration of renewables. While challenges remain, including intermittency and infrastructure development, the momentum behind renewable energy is undeniable, paving the way for a sustainable energy future.",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 3.2682197093963623,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "Renewable energy sources like solar, wind, and hydro are rapidly replacing fossil fuels due to technological advancements and supportive policies. Lower costs and increased reliability are driving a global shift towards cleaner, sustainable power, crucial for combating climate change."
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "multilingual",
    "test_name": "spanish_question",
    "prompt": "¿Qué es la inteligencia artificial y por qué es importante?",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 35.66495895385742,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "## ¿Qué es la Inteligencia Artificial (IA)?\n\nLa Inteligencia Artificial (IA) es un campo de la informática que se centra en diseñar y desarrollar sistemas que pueden realizar tareas que normalmente requieren inteligencia humana. En lugar de simplemente seguir instrucciones programadas, los sistemas de IA pueden aprender, razonar, resolver problemas, comprender el lenguaje natural y tomar decisiones.\n\nEn términos más específicos, la IA se puede dividir en varias ramas:\n\n* **Aprendizaje Automático (Machine Learning):**  Es la rama más popular actualmente.  Implica entrenar a sistemas para que aprendan de los datos sin ser programados explícitamente.  Ejemplos:  sistemas de recomendación (Netflix, Amazon), detección de spam, diagnóstico médico.\n* **Procesamiento del Lenguaje Natural (NLP):** Permite a las computadoras comprender, interpretar y generar lenguaje humano.  Ejemplos: chatbots, traducción automática, análisis de sentimientos.\n* **Visión Artificial (Computer Vision):**  Permite a las computadoras \"ver\" y comprender imágenes y videos. Ejemplos: reconocimiento facial, vehículos autónomos, diagnóstico médico por imágenes.\n* **Robótica:**  La IA se utiliza para desarrollar robots que pueden realizar tareas de forma autónoma.\n\n**En resumen, la IA busca crear máquinas que puedan pensar y actuar de manera similar a los humanos.**\n\n## ¿Por qué es importante la Inteligencia Artificial?\n\nLa IA es importante por una amplia gama de razones, con un impacto significativo en casi todos los aspectos de nuestra vida:\n\n**1. Beneficios Económicos:**\n\n* **Automatización:** La IA puede automatizar tareas repetitivas y tediosas, liberando a los humanos para tareas más creativas y estratégicas.\n* **Aumento de la Productividad:**  La IA puede optimizar procesos, mejorar la eficiencia y aumentar la producción en diversos sectores (manufactura, logística, finanzas).\n* **Nuevas Industrias:**  La IA está impulsando la creación de nuevas industrias y modelos de negocio.\n\n**2. Avances en la Salud:**\n\n* **Diagnóstico Médico:** La IA puede analizar imágenes médicas (radiografías, resonancias magnéticas) para detectar enfermedades con mayor precisión y rapidez.\n* **Descubrimiento de Fármacos:** La IA puede acelerar el proceso de descubrimiento y desarrollo de nuevos medicamentos.\n* **Atención al Paciente:**  Chatbots y asistentes virtuales pueden brindar atención"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "multilingual",
    "test_name": "french_question",
    "prompt": "Quels sont les avantages de l'apprentissage profond dans la vision par ordinateur?",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 35.82309865951538,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "L'apprentissage profond (Deep Learning) a révolutionné la vision par ordinateur, offrant des avantages considérables par rapport aux méthodes traditionnelles. Voici une liste des principaux avantages :\n\n**1. Performances Supérieures et Déjà Entraînées (Transfer Learning):**\n\n* **Meilleure précision:** Les réseaux neuronaux profonds, en particulier les réseaux convolutionnels (CNN), ont démontré une précision de pointe dans de nombreuses tâches de vision par ordinateur, dépassant souvent les méthodes traditionnelles comme les SVM ou les arbres de décision.\n* **Transfer Learning:** C'est l'un des avantages les plus significatifs. Grâce à l'apprentissage profond, on peut utiliser des modèles pré-entraînés sur de vastes ensembles de données (comme ImageNet) et les affiner pour une tâche spécifique. Cela réduit considérablement le temps de formation, les données et les ressources informatiques nécessaires.  On peut ainsi obtenir de bons résultats avec beaucoup moins de données et de temps de formation que si l'on devait entraîner un modèle à partir de zéro.\n\n**2. Automatisation et Réduction des Besoins Humains:**\n\n* **Apprentissage automatique:** L'apprentissage profond automatise le processus de construction de modèles d'apprentissage automatique, éliminant la nécessité d'une ingénierie manuelle complexe des caractéristiques (feature engineering).\n* **Détection d'objets:** Les réseaux neuronaux peuvent apprendre automatiquement des caractéristiques pertinentes des images, réduisant la nécessité d'une définition manuelle des caractéristiques.\n* **Segmentation sémantique:**  L'apprentissage profond peut segmenter automatiquement des images en différentes zones (par exemple, identifier les objets dans une image médicale)\n\n**3. Capacité à Traiter des Données Complexes:**\n\n* **Résolution des problèmes complexes:** L'apprentissage profond excelle dans la détection d'patterns complexes et subtils dans les images, ce qui est crucial pour des applications telles que la reconnaissance faciale, la détection d'anomalies et la vision 3D.\n* **Données non structurées:**  L'apprentissage profond est capable de traiter des données non structurées, comme les images et vidéos, qui sont difficiles à analyser avec les méthodes traditionnelles.\n* **Images haute résolution:** Les architectures comme le ResNet et le Vision Transformer sont capables de gérer des images de très haute résolution, ce qui est essentiel pour la reconnaissance d'objets à grande échelle.\n\n**4."
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "multilingual",
    "test_name": "chinese_question",
    "prompt": "在机器翻译中最大的挑战是什么？",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 35.38864850997925,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "机器翻译（MT）在不断进步，但仍然面临着许多巨大的挑战。以下是机器翻译中最大的挑战，可以大致分为几个类别：\n\n**1. 语言和文化差异：**\n\n* **歧义性 (Ambiguity):** 语言本身就存在歧义，机器翻译很难准确理解语言的含义，尤其是在不同语言之间。同一个词语或短语在不同的语言中可能有不同的解释。\n* **文化差异 (Cultural Context):** 翻译不仅仅是词汇的转换，更需要理解文化背景、习俗、价值观和幽默感。机器翻译往往难以捕捉到这些微妙的文化信息，导致翻译不准确或不合适。\n* **俚语、习语和方言 (Slang, Idioms, and Dialects):** 这些都是语言的固有特征，机器翻译很难准确翻译，因为它们通常不具有通用性。\n* **语境依赖性 (Context Dependency):**  语言的意义很大程度上取决于语境。机器翻译难以理解句子或段落的完整含义，需要依赖上下文来进行准确翻译。\n\n**2. 语义理解和推理：**\n\n* **深层语义 (Deep Semantic Understanding):**  机器翻译依赖于统计模型和模式匹配，但它们难以理解语言的深层含义，例如隐含的含义、讽刺、隐喻等。\n* **常识知识 (Common Sense Reasoning):**  人类拥有大量的常识知识，可以帮助机器翻译更好地理解文本的含义。然而，机器翻译缺乏这种常识知识，容易产生错误或不合理的结果。\n* **推理能力 (Reasoning Ability):** 机器翻译需要进行推理，才能理解文本的逻辑关系和隐含意义。  当前的翻译模型往往难以进行复杂的推理。\n\n**3. 文本结构和复杂性：**\n\n* **长文本处理 (Long-Text Processing):**  机器翻译在处理长文本时，更容易出现问题，例如句子之间的逻辑关系不清晰，文本结构混乱。\n* **结构化文本 (Structured Text):**  机器翻译对于结构化文本（如代码、法律文件、科学论文）的翻译尤其困难，因为它们需要理解文本的结构和关系。\n* **非线性文本 (Non-Linear Text):** 文本的结构和逻辑关系复杂，使得机器翻译难以准确理解和翻译。\n\n**4. 数据和模型限制：**\n\n* **训练数据的不足"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "multilingual",
    "test_name": "arabic_question",
    "prompt": "ما هي تطبيقات الذكاء الاصطناعي في مجال الرعاية الصحية؟",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 34.99179124832153,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "الذكاء الاصطناعي (AI) يغير بشكل كبير الطريقة التي يتم بها تقديم الرعاية الصحية، ويصل إلى مجموعة واسعة من التطبيقات. إليك بعض أبرز هذه التطبيقات، مصنفة حسب الفئة:\n\n**1. التشخيص الطبي:**\n\n* **تحليل الصور الطبية:** يُعتبر هذا من أكثر التطبيقات تطوراً. يمكن للذكاء الاصطناعي تحليل صور الأشعة (مثل الأشعة السينية، التصوير بالرنين المغناطيسي، التصوير المقطعي المحوسب) لتحديد التشوهات، مثل الأورام، أو الكسور، أو مشاكل القلب، بدقة أعلى من البشر في بعض الأحيان.\n* **تحليل بيانات المختبر:** يساعد الذكاء الاصطناعي في تحليل نتائج المختبرات، مثل تحليل عينات الدم للكشف عن الأمراض، وتحديد مستويات الأدوية، والتنبؤ بمسار المرضى.\n* **التشخيص الآلي:** تستخدم بعض التطبيقات الذكاء الاصطناعي لتحديد الأمراض بناءً على الأعراض والنتائج السريرية، مما يساعد الأطباء على اتخاذ قرارات أكثر دقة وفعالية.\n* **التنبؤ بالخطر:** يمكن للذكاء الاصطناعي تحليل البيانات السريرية للمرضى للتنبؤ بمدى خطر الإصابة بأمراض معينة، مما يسمح بالتدخل المبكر.\n\n**2. الأدوية والبحث:**\n\n* **اكتشاف الأدوية:** يساعد الذكاء الاصطناعي في تسريع عملية اكتشاف الأدوية من خلال تحليل كميات هائلة من البيانات الكيميائية والبيولوجية لتحديد الأهداف الدوائية المحتملة.\n* **التنبؤ بفعالية الأدوية:** يمكن للذكاء الاصطناعي تحليل البيانات السريرية للمرضى لتحديد ما إذا كان سيستجيبون لأحد الأدوية قبل البدء في العلاج.\n* **تخصيص العلاج:** يمكن للذكاء الاصطناعي تحليل البيانات الفردية للمرضى لتحديد العلاج الأنسب لهم، مع الأخذ في الاعتبار عوامل مثل التركيب الجيني والبيئة.\n\n**3. الرعاية الصحية الشخصية:**\n\n* **روبوتات الدردشة (Chatbots):** تُستخدم روبوتات الدردشة المدعومة بالذكاء الاصطناعي للإجابة على"
  },
  {
    "model_id": "google/gemma-3-1b-it",
    "category": "multilingual",
    "test_name": "german_question",
    "prompt": "Wie unterscheidet sich überwachtes Lernen vom unüberwachten Lernen?",
    "load_time_s": 5.7008116245269775,
    "inference_time_s": 35.01560044288635,
    "gpu_after_load_mb": 1907.15771484375,
    "output": "Okay, lass uns die Unterschiede zwischen überwachtem und unüberwachtem Lernen aufschlüsseln. Sie sind zwei grundlegend unterschiedliche Ansätze im maschinellen Lernen, die sich in der Art und Weise unterscheiden, wie sie Daten verwenden, um Modelle zu erstellen.\n\n**Überwachtes Lernen**\n\n* **Was es ist:** Bei überwachtem Lernen wird ein Modell mit einem *gelabelten* Datensatz trainiert.  \"Gelabelt\" bedeutet, dass für jede Eingabe (Eigenschaft oder Merkmal) eine korrekte Ausgabe (Label) zugeordnet ist.\n* **Wie es funktioniert:** Das Modell lernt, die Beziehung zwischen den Eingaben und den Ausgaben zu erkennen. Es passt seine Parameter so an, dass es die korrekten Ausgaben für die gelabelten Eingaben vorhersagt.\n* **Beispiele:**\n    * **Klassifizierung:**  Vorhersage der Kategorie einer Eingabe (z. B. \"Katze\" oder \"Hund\" anhand eines Bildes).\n    * **Regression:** Vorhersage eines kontinuierlichen Wertes (z. B. die Hauspreise basierend auf Größe und Lage).\n* **Anwendungsfälle:**  Bilderkennung, Sprachmodellierung, Spam-Erkennung, medizinische Diagnose, Kreditrisikobewertung.\n* **Wichtige Begriffe:**  Label-Daten, Trainingsdaten, Vorhersage, Fehler (die das Modell anhand der gelabelten Daten lernt).\n\n**Unüberwachtes Lernen**\n\n* **Was es ist:** Bei unüberwachtem Lernen wird ein Datensatz ohne gelabelte Daten verwendet.  Das Ziel ist es, Muster, Strukturen und Beziehungen in den Daten selbst zu entdecken.\n* **Wie es funktioniert:**  Das Modell versucht, die Daten selbst zu gruppieren, zu strukturieren oder zu reduzieren.  Es findet \"Cluster\" oder \"Ähnlichkeiten\" in den Daten.\n* **Beispiele:**\n    * **Clustering:** Gruppierung ähnlicher Datenpunkte (z. B. Kundensegmentierung).\n    * **Dimensionsreduktion:** Reduzierung der Anzahl der Variablen, während wichtige Informationen erhalten bleiben.  (z.B. Principal Component Analysis - PCA)\n    * **Assoziationsanalyse:**  Finden von Beziehungen zwischen Variablen (z. B. \"Menschen, die Eis essen, kaufen auch Schokolade\").\n* **An"
  }
]

import os
import time
import json
import torch
import argparse
import requests
from PIL import Image
from io import BytesIO
from datetime import datetime
from dotenv import load_dotenv
from transformers import AutoProcessor, Gemma3ForConditionalGeneration
from huggingface_hub import login
import gc

# Load environment variables from .env file
load_dotenv()

# Get Hugging Face token from environment variables
HF_TOKEN = os.getenv("HF_TOKEN")
if not HF_TOKEN:
    raise ValueError("HF_TOKEN not found in environment variables. Please check your .env file.")

# Login to Hugging Face Hub using the token
login(token=HF_TOKEN)
print("Successfully logged in to Hugging Face Hub")

# A sample image to use for multimodal models
SAMPLE_IMAGE_URL = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"

# Test prompts with images - using the same prompts but adding an image
TEST_PROMPTS = {
    "basic_qa": [
        {
            "name": "transfer_learning",
            "prompt": "Explain the concept of transfer learning in a concise paragraph."
        },
        {
            "name": "ada_lovelace",
            "prompt": "Who was Ada Lovelace, and why is she significant in computing history?"
        },
        {
            "name": "overfitting_definition",
            "prompt": "What does the term 'overfitting' mean in machine learning?"
        },
        {
            "name": "gradient_descent",
            "prompt": "Briefly describe gradient descent and how it works."
        },
        {
            "name": "supervised_vs_unsupervised",
            "prompt": "Name two major differences between supervised and unsupervised learning."
        }
    ],
    "code_generation": [
        {
            "name": "sum_even_numbers",
            "prompt": "Write a Python function that takes a list of integers and returns the sum of all even numbers."
        },
        {
            "name": "debug_snippet",
            "prompt": (
                "Here is a Python snippet that's causing an error. Please fix it:\n\n"
                "```python\ndef greet(name):\n print(Hello, name)\n```"
            )
        },
        {
            "name": "factorial_function",
            "prompt": "Create a Python function that calculates the factorial of a given integer."
        },
        {
            "name": "cpp_hello_world",
            "prompt": "Generate a minimal C++ code snippet that prints 'Hello World' to the console."
        },
        {
            "name": "optimize_loops",
            "prompt": (
                "Optimize this Python code for performance:\n\n"
                "```python\n"
                "for i in range(10000):\n"
                " for j in range(10000):\n"
                " pass\n"
                "```"
            )
        }
    ],
    "summarization": [
        {
            "name": "internet_history",
            "prompt": (
                "Summarize the following text in under 50 words:\n\n"
                "The Internet's origins trace back to the 1960s, during the Cold War, when the U.S. Department of Defense's "
                "Advanced Research Projects Agency (ARPA) funded ARPANET, a project designed to create a decentralized "
                "communication network that could withstand a nuclear attack. In the 1970s, TCP/IP protocols were developed, "
                "laying the foundation for the modern Internet. The 1980s saw the rise of the Domain Name System (DNS) and "
                "the National Science Foundation Network (NSFNET), which expanded access beyond military and academic "
                "institutions. The World Wide Web, invented by Tim Berners-Lee in 1989, revolutionized the Internet by "
                "introducing hypertext and a user-friendly interface. The 1990s marked the commercialization of the Internet, "
                "with the emergence of web browsers and e-commerce. Today, the Internet is a global network connecting billions "
                "of devices and shaping nearly every aspect of modern life."
            )
        },
        {
            "name": "climate_change_impact",
            "prompt": (
                "Summarize the following text in under 50 words:\n\n"
                "Climate change, driven by the increasing concentration of greenhouse gases in the atmosphere, is causing "
                "significant and widespread impacts on the planet. Rising global temperatures lead to melting polar ice caps "
                "and glaciers, resulting in sea-level rise and coastal flooding. Extreme weather events, such as hurricanes, "
                "droughts, and wildfires, are becoming more frequent and intense. Changes in precipitation patterns disrupt "
                "agriculture and water resources, threatening food security. Ocean acidification, caused by increased "
                "absorption of carbon dioxide, harms marine ecosystems. These impacts have profound consequences for human "
                "societies, economies, and natural environments, necessitating urgent action to mitigate and adapt to climate "
                "change."
            )
        },
        {
            "name": "ai_healthcare",
            "prompt": (
                "Summarize the following text in under 50 words:\n\n"
                "Artificial intelligence (AI) is transforming healthcare by enabling faster and more accurate diagnoses, "
                "personalized treatments, and improved patient outcomes. AI-powered algorithms can analyze medical images, "
                "such as X-rays and MRIs, to detect diseases like cancer and Alzheimer's earlier than human experts. Machine "
                "learning models can predict patient risk for various conditions, allowing for proactive interventions. AI-driven "
                "drug discovery accelerates the development of new therapies. Virtual assistants and chatbots provide patients "
                "with 24/7 access to medical information and support. While AI offers immense potential, challenges remain, "
                "including data privacy concerns, regulatory hurdles, and the need for seamless integration into existing "
                "healthcare systems."
            )
        },
        {
            "name": "mobile_evolution",
            "prompt": (
                "Summarize the following text in under 50 words:\n\n"
                "The evolution of mobile technology has dramatically reshaped communication and information access. From bulky "
                "analog phones in the 1980s, mobile devices have transformed into powerful handheld computers. The introduction "
                "of 2G networks enabled text messaging and basic data services, while 3G brought faster internet speeds and "
                "multimedia capabilities. The advent of 4G LTE revolutionized mobile broadband, supporting high-definition "
                "video streaming and real-time applications. 5G technology promises even faster speeds, lower latency, and "
                "greater network capacity, enabling new applications like augmented reality and autonomous vehicles. The "
                "proliferation of smartphones and mobile apps has created a ubiquitous computing environment, impacting "
                "everything from social interactions to business operations."
            )
        },
        {
            "name": "renewable_energy",
            "prompt": (
                "Summarize the following text in under 50 words:\n\n"
                "The rise of renewable energy sources is a critical component of the global effort to combat climate change. "
                "Solar, wind, hydro, and geothermal power are increasingly replacing fossil fuels, reducing greenhouse gas "
                "emissions and air pollution. Technological advancements have driven down the cost of renewable energy, "
                "making it economically competitive with traditional sources. Government policies, such as subsidies and "
                "carbon pricing, are accelerating the transition to clean energy. Energy storage solutions, like batteries, "
                "are improving the reliability and grid integration of renewables. While challenges remain, including "
                "intermittency and infrastructure development, the momentum behind renewable energy is undeniable, paving the "
                "way for a sustainable energy future."
            )
        }
    ],
    "multilingual": [
        {
            "name": "spanish_question",
            "prompt": "¿Qué es la inteligencia artificial y por qué es importante?"
        },
        {
            "name": "french_question",
            "prompt": "Quels sont les avantages de l'apprentissage profond dans la vision par ordinateur?"
        },
        {
            "name": "chinese_question",
            "prompt": "在机器翻译中最大的挑战是什么？"
        },
        {
            "name": "arabic_question",
            "prompt": "ما هي تطبيقات الذكاء الاصطناعي في مجال الرعاية الصحية؟"
        },
        {
            "name": "german_question",
            "prompt": "Wie unterscheidet sich überwachtes Lernen vom unüberwachten Lernen?"
        }
    ]
}

def log_gpu_memory():
    """Log and return GPU memory usage in MB."""
    if torch.cuda.is_available():
        devices = list(range(torch.cuda.device_count()))
        memory_allocated = []
        
        for device in devices:
            mem_allocated = torch.cuda.memory_allocated(device) / (1024 ** 2)
            mem_reserved = torch.cuda.memory_reserved(device) / (1024 ** 2)
            memory_allocated.append(mem_allocated)
            print(f"GPU {device}: Allocated: {mem_allocated:.2f} MB, Reserved: {mem_reserved:.2f} MB")
        
        return sum(memory_allocated)
    else:
        print("No GPU available")
        return 0

def clear_gpu_memory():
    """Clear GPU memory cache."""
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        print("GPU memory cache cleared")
        gc.collect()
        return log_gpu_memory()
    else:
        print("No GPU available")
        return 0

def get_sample_image():
    """Download a sample image for the multimodal tests."""
    response = requests.get(SAMPLE_IMAGE_URL)
    return Image.open(BytesIO(response.content))

def format_message(prompt, image):
    """Format a text and image prompt as a message for the model."""
    return [
        {
            "role": "system",
            "content": [{"type": "text", "text": "You are a helpful assistant."}]
        },
        {
            "role": "user",
            "content": [
                {"type": "image", "image": image},
                {"type": "text", "text": prompt}
            ]
        }
    ]

def test_model(model_id, output_dir="./results"):
    """Test a multimodal model using direct model loading."""
    print(f"\n{'='*50}\nTESTING MODEL: {model_id}\n{'='*50}")
    
    # Create output directory
    os.makedirs(output_dir, exist_ok=True)
    
    # Create results file with timestamp
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    model_name = model_id.split("/")[-1]
    result_file = os.path.join(output_dir, f"{model_name}_{timestamp}.json")
    results = []
    
    try:
        # Log initial GPU state
        print("Initial GPU memory state:")
        initial_gpu_mem = log_gpu_memory()
        
        # Load model and processor
        print(f"Loading model: {model_id}")
        start_time = time.time()
        
        print("Device set to use cuda:0")
        device = "cuda:0" if torch.cuda.is_available() else "cpu"
        
        # Download a sample image
        print("Loading sample image...")
        sample_image = get_sample_image()
        
        # Load the model and processor - FOR MULTIMODAL MODELS
        processor = AutoProcessor.from_pretrained(model_id, token=HF_TOKEN)
        model = Gemma3ForConditionalGeneration.from_pretrained(
            model_id, 
            device_map="auto",
            torch_dtype=torch.bfloat16,
            token=HF_TOKEN
        )
        
        load_time = time.time() - start_time
        print(f"Model loaded in {load_time:.2f} seconds")
        
        # Log GPU memory after loading
        print("GPU memory after model load:")
        gpu_after_load = log_gpu_memory()
        
        # Run tests for each category and prompt
        for category, prompts in TEST_PROMPTS.items():
            print(f"\n--- Testing {category} prompts ---")
            
            for prompt_data in prompts:
                prompt_name = prompt_data["name"]
                prompt_text = prompt_data["prompt"]
                
                print(f"Running: {prompt_name}")
                
                try:
                    # Start timing
                    inference_start = time.time()
                    
                    # Format as a chat message with the image for multimodal models
                    messages = format_message(prompt_text, sample_image)
                    
                    # Apply chat template to format for model
                    inputs = processor.apply_chat_template(
                        messages, 
                        add_generation_prompt=True, 
                        tokenize=True,
                        return_dict=True, 
                        return_tensors="pt"
                    ).to(model.device, dtype=torch.bfloat16)
                    
                    # Remember input length to extract only the response later
                    input_len = inputs["input_ids"].shape[-1]
                    
                    # Run inference
                    with torch.inference_mode():
                        generation = model.generate(
                            **inputs, 
                            max_new_tokens=512, 
                            do_sample=True,
                            temperature=0.7,
                            top_p=0.9
                        )
                        generation = generation[0][input_len:]
                    
                    # Decode to get the text
                    response = processor.decode(generation, skip_special_tokens=True)
                    
                    inference_time = time.time() - inference_start
                    print(f"Completed in {inference_time:.2f} seconds")
                    
                except Exception as e:
                    print(f"Error during inference: {str(e)}")
                    import traceback
                    traceback.print_exc()
                    response = f"ERROR: {str(e)}"
                    inference_time = time.time() - inference_start
                
                # Record results
                result_entry = {
                    "model_id": model_id,
                    "category": category,
                    "test_name": prompt_name,
                    "prompt": prompt_text,
                    "load_time_s": load_time,
                    "inference_time_s": inference_time,
                    "gpu_after_load_mb": gpu_after_load,
                    "output": response
                }
                
                results.append(result_entry)
                
                # Save results after each prompt
                with open(result_file, 'w', encoding='utf-8') as f:
                    json.dump(results, f, indent=2, ensure_ascii=False)
        
        print(f"\nAll tests completed for {model_id}. Results saved to: {result_file}")
        
    except Exception as e:
        print(f"Error testing model {model_id}: {str(e)}")
        import traceback
        traceback.print_exc()
    
    finally:
        # Clean up
        if 'model' in locals():
            del model
        if 'processor' in locals():
            del processor
        gc.collect()
        clear_gpu_memory()
    
    return result_file

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Test Gemma 3 multimodal models")
    parser.add_argument(
        "--model", 
        type=str,
        default="google/gemma-3-4b-it",
        help="Multimodal HuggingFace model ID to test"
    )
    parser.add_argument(
        "--output_dir", 
        type=str, 
        default="./results",
        help="Directory to save results"
    )
    
    args = parser.parse_args()
    
    # Check CUDA availability
    if torch.cuda.is_available():
        print(f"CUDA available: {torch.cuda.is_available()}")
        print(f"CUDA version: {torch.version.cuda}")
        print(f"Number of GPUs: {torch.cuda.device_count()}")
        for i in range(torch.cuda.device_count()):
            print(f"GPU {i}: {torch.cuda.get_device_name(i)}")
    else:
        print("WARNING: CUDA not available. Tests will run on CPU.")
    
    # Test the specified multimodal model
    test_model(args.model, args.output_dir)

[
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "basic_qa",
    "test_name": "transfer_learning",
    "prompt": "Explain the concept of transfer learning in a concise paragraph.",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 16.354753971099854,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Okay, here’s a concise explanation of transfer learning:\n\nTransfer learning is a machine learning technique where a model trained on one task is re-purposed as the starting point for a model on a second, related task. Instead of training a new model from scratch, you leverage the knowledge (learned features and patterns) gained from the initial task. This significantly reduces training time, improves performance, and often requires less data for the new task, especially when the datasets are small. \n\n---\n\nWould you like me to elaborate on any specific aspect of transfer learning, or perhaps give an example of how it’s used?"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "basic_qa",
    "test_name": "ada_lovelace",
    "prompt": "Who was Ada Lovelace, and why is she significant in computing history?",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 47.79332971572876,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Okay, let's dive into the fascinating story of Ada Lovelace!\n\n**Who was Ada Lovelace?**\n\nAda Lovelace (1815-1852) was an English mathematician and writer. She's primarily known for her work on Charles Babbage’s proposed Analytical Engine, a mechanical general-purpose computer. She was the daughter of Lord Byron, the famous poet, and Annabella Milbanke, a mathematician. \n\n**Why is she significant in computing history?**\n\nHere's what makes Ada Lovelace so important:\n\n* **First Computer Programmer:**  This is the most widely recognized aspect of her legacy. In 1843, she translated an article by Italian military engineer Luigi Menabrea about Babbage's Analytical Engine. However, she didn't just translate it; she added extensive notes of her own—nearly three times the length of the original article. \n\n* **Algorithmic Thinking:** Within those notes, specifically Note G, she described an algorithm for the Analytical Engine to calculate a sequence of Bernoulli numbers. This is considered the *first* algorithm intended to be processed by a machine. It’s a groundbreaking concept – she essentially envisioned a machine performing a specific task based on a set of instructions. \n\n* **Understanding Beyond Calculation:** Crucially, Lovelace went *beyond* simply seeing the Analytical Engine as a glorified calculator. She recognized its potential to do far more than just crunch numbers. She wrote about the possibility of the machine composing elaborate pieces of music or producing graphics – essentially, she foresaw the concept of general-purpose computing. She wrote: \"The Analytical Engine has no pretensions whatever to *originate* anything. It can do whatever we *know how to order* it to perform.\"  This was a remarkably insightful statement about the limitations and potential of machines.\n\n* **A Pioneer Ahead of Her Time:**  The Analytical Engine was never actually built during her lifetime (Babbage struggled with funding and engineering challenges).  But Lovelace’s notes demonstrated a deep understanding of the machine’s possibilities and laid the conceptual groundwork for modern computing.\n\n**In short, Ada Lovelace is celebrated as the first computer programmer because she conceived of and described an algorithm – a set of instructions – for a machine to execute, a concept fundamental to the field of computer science.**\n\n---\n\nWould you like me to:\n\n*   Tell you more about the Analytical Engine?\n*   Discuss other early pioneers of computing"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "basic_qa",
    "test_name": "overfitting_definition",
    "prompt": "What does the term 'overfitting' mean in machine learning?",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 46.8962984085083,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Okay, let's break down the concept of \"overfitting\" in machine learning.\n\n**What is Overfitting?**\n\nIn simple terms, overfitting happens when a machine learning model learns the training data *too* well. It essentially memorizes the training data, including its noise and specific quirks, instead of learning the underlying patterns that would allow it to generalize to new, unseen data.\n\n**Here's a more detailed explanation:**\n\n1. **Training Data:** You feed your model a dataset of examples (e.g., images of cats and dogs).\n\n2. **Model Learning:** The model analyzes this data and tries to find relationships between the features (e.g., color, shape, texture) and the labels (e.g., \"cat\" or \"dog\").\n\n3. **Overfitting Occurs:** If the model is too complex (e.g., too many layers in a neural network, or a very high-degree polynomial in a regression model), it can start to fit the training data *perfectly*.  It might even learn things that are just random fluctuations in the training data – things that aren’t actually meaningful features of cats or dogs. \n\n4. **Poor Generalization:** The problem is, when you give the model new, unseen data (e.g., a picture of a cat it’s never seen before), it performs poorly. It's unable to accurately classify it because it’s focused on the specific details of the training set rather than the general characteristics of cats.\n\n\n\n**Think of it like this:**\n\nImagine you're teaching a child to recognize cats.\n\n*   **Good Learning:** You show them pictures of various cats – different breeds, colors, poses. They learn to identify *what makes a cat a cat* – pointy ears, whiskers, a tail, etc. They can then recognize a new cat they've never seen before.\n\n*   **Overfitting:** You only show them *one* specific fluffy, gray tabby cat. The child might then only be able to recognize *that specific cat* and not other cats.\n\n\n\n**How to Detect and Prevent Overfitting:**\n\n*   **More Data:** Often, the simplest solution is to get more training data.\n*   **Simplify the Model:** Use a less complex model (e.g., fewer layers in a neural network, a lower-degree polynomial).\n*   **Regularization:** Techniques like L1 or L2 regularization add a penalty"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "basic_qa",
    "test_name": "gradient_descent",
    "prompt": "Briefly describe gradient descent and how it works.",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 34.32464265823364,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Okay, let's break down gradient descent.\n\n**What is Gradient Descent?**\n\nGradient descent is an optimization algorithm used to find the minimum of a function.  In machine learning, this function is often a *cost function* – it measures how \"wrong\" a model's predictions are.  The goal is to adjust the model's parameters to minimize this cost, making the model more accurate.\n\n**How it Works:**\n\n1. **Start with a Guess:** You begin with an initial guess for the values of the model's parameters.\n\n2. **Calculate the Gradient:** The \"gradient\" is a vector that points in the direction of the *steepest increase* of the cost function.  Think of it like the slope of a hill.\n\n3. **Move in the Opposite Direction:**  Since we want to *minimize* the cost, we take a step in the *opposite* direction of the gradient. This is like rolling downhill.\n\n4. **Adjust Parameters:**  We update the model's parameters based on this step.  The size of the step is determined by a \"learning rate.\"\n\n5. **Repeat:** Steps 2-4 are repeated iteratively.  With each iteration, the algorithm moves closer to the minimum of the cost function.\n\n\n**Analogy:**\n\nImagine you're blindfolded on a hilly landscape and trying to find the lowest point. You can feel the slope of the ground beneath your feet. Gradient descent is like taking small steps downhill in the direction where the ground is steepest.\n\n\n\n**Do you want me to delve deeper into any specific aspect of gradient descent, such as:**\n\n*   Learning rate?\n*   Different types of gradient descent (e.g., stochastic gradient descent)?\n*   How it relates to neural networks?"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "basic_qa",
    "test_name": "supervised_vs_unsupervised",
    "prompt": "Name two major differences between supervised and unsupervised learning.",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 27.96626043319702,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Okay, let’s break down the differences between supervised and unsupervised learning. Here are two major distinctions:\n\n1.  **Labeled Data:**\n    *   **Supervised Learning:**  Requires *labeled* data. This means each data point has an associated “correct answer” or target variable. For example, you might have images of cats and dogs, each labeled as either \"cat\" or \"dog.\" The algorithm learns to map the input (the image) to the output (the label).\n    *   **Unsupervised Learning:** Works with *unlabeled* data. There are no pre-defined answers or target variables. The algorithm's job is to find patterns, structures, or groupings within the data itself.  For instance, you might give it a collection of customer purchase histories and it would try to identify different customer segments without knowing what those segments *are* beforehand.\n\n2.  **Goal of the Algorithm:**\n    *   **Supervised Learning:** The goal is to learn a function that can accurately predict the output for *new*, unseen data based on the patterns it learned from the labeled training data.\n    *   **Unsupervised Learning:** The goal is to discover hidden structures or relationships in the data. This could involve clustering similar data points together, reducing the dimensionality of the data, or identifying anomalies.\n\n---\n\nDo you want me to elaborate on any of these points, or perhaps discuss specific examples of algorithms used in each type of learning?"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "code_generation",
    "test_name": "sum_even_numbers",
    "prompt": "Write a Python function that takes a list of integers and returns the sum of all even numbers.",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 46.84884452819824,
    "gpu_after_load_mb": 8201.845703125,
    "output": "```python\ndef sum_even_numbers(numbers):\n  \"\"\"\n  Calculates the sum of all even numbers in a list.\n\n  Args:\n    numbers: A list of integers.\n\n  Returns:\n    The sum of all even numbers in the list.  Returns 0 if the list is empty or contains no even numbers.\n  \"\"\"\n  sum_of_evens = 0\n  for number in numbers:\n    if number % 2 == 0:\n      sum_of_evens += number\n  return sum_of_evens\n\n# Example usage:\nnumbers1 = [1, 2, 3, 4, 5, 6]\nresult1 = sum_even_numbers(numbers1)\nprint(f\"The sum of even numbers in {numbers1} is: {result1}\")  # Output: 12\n\nnumbers2 = [1, 3, 5, 7]\nresult2 = sum_even_numbers(numbers2)\nprint(f\"The sum of even numbers in {numbers2} is: {result2}\")  # Output: 0\n\nnumbers3 = []\nresult3 = sum_even_numbers(numbers3)\nprint(f\"The sum of even numbers in {numbers3} is: {result3}\")  # Output: 0\n```\n\n**Explanation:**\n\n1. **Function Definition:**\n   - The code defines a function called `sum_even_numbers` that takes one argument: `numbers` (a list of integers).\n\n2. **Initialization:**\n   - `sum_of_evens = 0`:  A variable `sum_of_evens` is initialized to 0. This variable will store the sum of the even numbers found in the list.\n\n3. **Iteration:**\n   - `for number in numbers:`: The code iterates through each `number` in the input `numbers` list.\n\n4. **Even Number Check:**\n   - `if number % 2 == 0:`:  Inside the loop, the code checks if the current `number` is even.  The modulo operator (`%`) gives the remainder of a division. If the remainder when `number` is divided by 2 is 0, then the number is even.\n\n5. **Summation:**\n   - `sum_of_evens"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "code_generation",
    "test_name": "debug_snippet",
    "prompt": "Here is a Python snippet that's causing an error. Please fix it:\n\n```python\ndef greet(name):\n print(Hello, name)\n```",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 12.990885019302368,
    "gpu_after_load_mb": 8201.845703125,
    "output": "```python\ndef greet(name):\n  print(\"Hello, \" + name)\n```\n\n**Reasoning:**\n\nThe original code had a syntax error in the `print` statement.  The string \"Hello\" was not properly enclosed in quotes, causing Python to misinterpret it as a variable name.  By adding double quotes around the string, we tell Python that it's a literal string, which is what we intended.  Additionally, to concatenate the string \"Hello\" with the variable `name`, we need to use the `+` operator.  The corrected code concatenates the string \"Hello, \" with the value of the `name` variable."
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "code_generation",
    "test_name": "factorial_function",
    "prompt": "Create a Python function that calculates the factorial of a given integer.",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 46.90397024154663,
    "gpu_after_load_mb": 8201.845703125,
    "output": "```python\ndef factorial(n):\n  \"\"\"\n  Calculates the factorial of a non-negative integer.\n\n  Args:\n    n: A non-negative integer.\n\n  Returns:\n    The factorial of n (n!), or 1 if n is 0.\n    Returns an error message if n is negative.\n  \"\"\"\n  if n < 0:\n    return \"Factorial is not defined for negative numbers.\"\n  elif n == 0:\n    return 1\n  else:\n    result = 1\n    for i in range(1, n + 1):\n      result *= i\n    return result\n\n# Example usage:\nprint(factorial(5))  # Output: 120\nprint(factorial(0))  # Output: 1\nprint(factorial(-1)) # Output: Factorial is not defined for negative numbers.\n```\n\n**Explanation:**\n\n1. **Function Definition:**\n   - The code defines a function called `factorial(n)` that takes a single argument `n`, which represents the integer for which we want to calculate the factorial.\n\n2. **Error Handling:**\n   - It first checks if `n` is negative. If it is, it returns a message indicating that the factorial is not defined for negative numbers.\n\n3. **Base Case:**\n   - If `n` is 0, it returns 1 because the factorial of 0 is defined as 1.\n\n4. **Iterative Calculation:**\n   - If `n` is a positive integer, it initializes a variable `result` to 1.\n   - It then uses a `for` loop to iterate from 1 to `n` (inclusive).\n   - In each iteration, it multiplies `result` by the current value of `i`. This effectively calculates the product of all positive integers from 1 to `n`.\n\n5. **Return Value:**\n   - Finally, it returns the calculated `result`, which is the factorial of `n`.\n\n**How the factorial calculation works:**\n\nThe factorial of a non-negative integer `n`, denoted as `n!`, is the product of all positive integers less than or equal to `n`.  For example:\n\n- 5! = 5 * 4 * 3 * 2 * 1 = 120\n- 0! = 1 (by definition"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "code_generation",
    "test_name": "cpp_hello_world",
    "prompt": "Generate a minimal C++ code snippet that prints 'Hello World' to the console.",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 4.009864330291748,
    "gpu_after_load_mb": 8201.845703125,
    "output": "```cpp\n#include <iostream>\n\nint main() {\n  std::cout << \"Hello World\" << std::endl;\n  return 0;\n}\n```\n"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "code_generation",
    "test_name": "optimize_loops",
    "prompt": "Optimize this Python code for performance:\n\n```python\nfor i in range(10000):\n for j in range(10000):\n pass\n```",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 46.36426663398743,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Okay, let's optimize that Python code. The code you've provided is extremely basic and doesn't really *need* optimization because it's just a nested loop that does nothing. However, I'll demonstrate how to make it more efficient and explain why it's beneficial.\n\n**Original Code:**\n\n```python\nfor i in range(10000):\n for j in range(10000):\n pass\n```\n\n**Optimized Code (No Change Needed - But with Explanation):**\n\n```python\n# The original code is already as efficient as it can be for its purpose.\n# It's just a placeholder. If you were doing something more meaningful\n# within the loops, then optimization would be relevant.\nfor i in range(10000):\n for j in range(10000):\n pass\n```\n\n**Explanation and Why It's Already \"Optimized\"**\n\n1. **`pass` Statement:**  The `pass` statement does nothing. It's a placeholder that is required syntactically in many situations where a statement is expected but you don't want any code to execute.  In this case, it's perfectly acceptable.\n\n2. **Nested Loops:** Nested loops are often a performance bottleneck in Python.  The time complexity of this code is O(n*m), where n and m are 10000 in this case.  This means the execution time grows quadratically with the input size.\n\n3. **No Optimization Possible (Without Changing the Purpose):** Since the code does nothing, there's no way to make it faster *without changing its purpose*. If the goal is to simply iterate through two loops, the code is already as efficient as it can be.\n\n**When Optimization *Would* Be Relevant**\n\nIf you were doing something *inside* those loops (e.g., calculations, data manipulation, comparisons), then optimization would be important. Here are some common optimization techniques:\n\n* **Avoid Loops When Possible:**  Python's built-in functions and libraries (like `map`, `filter`, `reduce`, list comprehensions, NumPy) are often much faster than explicit loops, especially for numerical operations.\n\n* **Vectorization (NumPy):** If you're working with numerical data, NumPy can significantly speed up calculations by operating on entire arrays at once rather than looping through individual elements.\n\n*"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "summarization",
    "test_name": "internet_history",
    "prompt": "Summarize the following text in under 50 words:\n\nThe Internet's origins trace back to the 1960s, during the Cold War, when the U.S. Department of Defense's Advanced Research Projects Agency (ARPA) funded ARPANET, a project designed to create a decentralized communication network that could withstand a nuclear attack. In the 1970s, TCP/IP protocols were developed, laying the foundation for the modern Internet. The 1980s saw the rise of the Domain Name System (DNS) and the National Science Foundation Network (NSFNET), which expanded access beyond military and academic institutions. The World Wide Web, invented by Tim Berners-Lee in 1989, revolutionized the Internet by introducing hypertext and a user-friendly interface. The 1990s marked the commercialization of the Internet, with the emergence of web browsers and e-commerce. Today, the Internet is a global network connecting billions of devices and shaping nearly every aspect of modern life.",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 7.806931018829346,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Here’s a summary of the text in under 50 words:\n\nThe Internet began in the 1960s as ARPANET, a Cold War project for resilient communication. Developments like TCP/IP and DNS expanded its reach, culminating in the World Wide Web’s invention in 1989 and the Internet’s commercialization in the 90s."
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "summarization",
    "test_name": "climate_change_impact",
    "prompt": "Summarize the following text in under 50 words:\n\nClimate change, driven by the increasing concentration of greenhouse gases in the atmosphere, is causing significant and widespread impacts on the planet. Rising global temperatures lead to melting polar ice caps and glaciers, resulting in sea-level rise and coastal flooding. Extreme weather events, such as hurricanes, droughts, and wildfires, are becoming more frequent and intense. Changes in precipitation patterns disrupt agriculture and water resources, threatening food security. Ocean acidification, caused by increased absorption of carbon dioxide, harms marine ecosystems. These impacts have profound consequences for human societies, economies, and natural environments, necessitating urgent action to mitigate and adapt to climate change.",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 5.906728506088257,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Here’s a summary of the text in under 50 words:\n\nClimate change, fueled by greenhouse gases, is causing rising temperatures, extreme weather, and rising sea levels. These changes threaten food security, marine life, and human societies, demanding immediate action to mitigate its devastating global impacts."
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "summarization",
    "test_name": "ai_healthcare",
    "prompt": "Summarize the following text in under 50 words:\n\nArtificial intelligence (AI) is transforming healthcare by enabling faster and more accurate diagnoses, personalized treatments, and improved patient outcomes. AI-powered algorithms can analyze medical images, such as X-rays and MRIs, to detect diseases like cancer and Alzheimer's earlier than human experts. Machine learning models can predict patient risk for various conditions, allowing for proactive interventions. AI-driven drug discovery accelerates the development of new therapies. Virtual assistants and chatbots provide patients with 24/7 access to medical information and support. While AI offers immense potential, challenges remain, including data privacy concerns, regulatory hurdles, and the need for seamless integration into existing healthcare systems.",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 6.20838737487793,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Here’s a summary of the text in under 50 words:\n\nAI is revolutionizing healthcare through faster diagnoses, personalized treatments, and proactive patient care. It aids in disease detection, drug development, and offers virtual support. However, challenges like data privacy and system integration must be addressed for widespread adoption."
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "summarization",
    "test_name": "mobile_evolution",
    "prompt": "Summarize the following text in under 50 words:\n\nThe evolution of mobile technology has dramatically reshaped communication and information access. From bulky analog phones in the 1980s, mobile devices have transformed into powerful handheld computers. The introduction of 2G networks enabled text messaging and basic data services, while 3G brought faster internet speeds and multimedia capabilities. The advent of 4G LTE revolutionized mobile broadband, supporting high-definition video streaming and real-time applications. 5G technology promises even faster speeds, lower latency, and greater network capacity, enabling new applications like augmented reality and autonomous vehicles. The proliferation of smartphones and mobile apps has created a ubiquitous computing environment, impacting everything from social interactions to business operations.",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 6.748497009277344,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Here’s a summary of the text in under 50 words:\n\nMobile technology has evolved rapidly, from basic phones to powerful smartphones. Each generation – 2G, 3G, 4G, and now 5G – has increased speed and capabilities, transforming communication, information access, and creating a ubiquitous computing environment."
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "summarization",
    "test_name": "renewable_energy",
    "prompt": "Summarize the following text in under 50 words:\n\nThe rise of renewable energy sources is a critical component of the global effort to combat climate change. Solar, wind, hydro, and geothermal power are increasingly replacing fossil fuels, reducing greenhouse gas emissions and air pollution. Technological advancements have driven down the cost of renewable energy, making it economically competitive with traditional sources. Government policies, such as subsidies and carbon pricing, are accelerating the transition to clean energy. Energy storage solutions, like batteries, are improving the reliability and grid integration of renewables. While challenges remain, including intermittency and infrastructure development, the momentum behind renewable energy is undeniable, paving the way for a sustainable energy future.",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 6.031820058822632,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Here’s a summary of the text in under 50 words:\n\nRenewable energy sources – solar, wind, and hydro – are rapidly replacing fossil fuels due to falling costs and supportive policies. Technological advancements and energy storage are boosting their reliability, driving a sustainable energy future and reducing harmful emissions."
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "multilingual",
    "test_name": "spanish_question",
    "prompt": "¿Qué es la inteligencia artificial y por qué es importante?",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 47.038649559020996,
    "gpu_after_load_mb": 8201.845703125,
    "output": "¡Claro! Con gusto te explico qué es la inteligencia artificial (IA) y por qué es tan importante.\n\n**¿Qué es la Inteligencia Artificial?**\n\nEn términos sencillos, la inteligencia artificial es la capacidad de una computadora o máquina para imitar funciones cognitivas humanas, como:\n\n*   **Aprender:**  La IA puede analizar datos y mejorar su rendimiento con el tiempo sin ser programada explícitamente para cada situación.\n*   **Razonar:**  Puede resolver problemas y tomar decisiones basadas en la información que tiene.\n*   **Percibir:**  Puede \"ver\" (a través de la visión artificial), \"oír\" (a través del procesamiento del lenguaje natural) y \"sentir\" (a través de sensores).\n*   **Comprender el lenguaje:**  Puede entender y responder al lenguaje humano, como en los chatbots.\n\n**Existen diferentes tipos de IA:**\n\n*   **IA Débil o Estrecha:**  Está diseñada para realizar una tarea específica, como jugar al ajedrez, recomendar productos o reconocer rostros. Es la IA que más vemos en uso actualmente.\n*   **IA Fuerte o General:**  Es un tipo de IA hipotético que tendría la capacidad de comprender, aprender y aplicar su inteligencia a cualquier tarea que un humano pueda realizar.\n*   **Superinteligencia:**  Un concepto aún más teórico, donde la IA superaría la inteligencia humana en todos los aspectos.\n\n**¿Por qué es importante la IA?**\n\nLa IA está teniendo un impacto cada vez mayor en casi todos los aspectos de nuestras vidas y tiene el potencial de transformar el mundo de muchas maneras:\n\n*   **Automatización:**  La IA puede automatizar tareas repetitivas y peligrosas, liberando a los humanos para que se concentren en trabajos más creativos y estratégicos.\n*   **Mejora de la eficiencia:**  Puede optimizar procesos en industrias como la manufactura, la logística y la energía, reduciendo costos y mejorando la productividad.\n*   **Avances en la medicina:**  La IA se utiliza para diagnosticar enfermedades, desarrollar nuevos medicamentos y personalizar tratamientos.\n*   **Transporte:**  La conducción autónoma es una aplicación clave de la IA que podría revolucionar la forma en que nos movemos.\n*   **Atención al cliente:**  Los chatbots y asistentes virtuales impulsados por IA están mejorando la experiencia del cliente.\n"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "multilingual",
    "test_name": "french_question",
    "prompt": "Quels sont les avantages de l'apprentissage profond dans la vision par ordinateur?",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 45.22296905517578,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Bien sûr, voici quelques avantages de l'apprentissage profond dans la vision par ordinateur :\n\n*   **Représentation des caractéristiques automatique :** les modèles d'apprentissage profond, en particulier les réseaux neuronaux convolutifs (CNN), peuvent apprendre automatiquement des caractéristiques pertinentes à partir des données brutes, sans nécessiter d'ingénierie manuelle des caractéristiques. Cela réduit considérablement le temps et les efforts nécessaires pour développer des systèmes de vision par ordinateur.\n*   **Précision améliorée :** les modèles d'apprentissage profond ont atteint des performances de pointe dans diverses tâches de vision par ordinateur, telles que la classification d'images, la détection d'objets et la segmentation d'images. Ils surpassent souvent les méthodes traditionnelles en termes de précision et de robustesse.\n*   **Capacité de généralisation :** les modèles d'apprentissage profond peuvent généraliser à de nouvelles données qui ne sont pas vues pendant l'entraînement, ce qui leur permet de fonctionner bien dans des scénarios du monde réel.\n*   **Traitement des données complexes€:** l'apprentissage profond peut traiter des données complexes et non linéaires qui sont difficiles à modéliser avec des méthodes traditionnelles.\n*   **Adaptabilité :** les modèles d'apprentissage profond peuvent être adaptés à une grande variété de tâches de vision par ordinateur en modifiant l'architecture du réseau ou en affinant les paramètres sur des données spécifiques.\n\nVoici quelques exemples d'applications de l'apprentissage profond dans la vision par ordinateur :\n\n*   **Voitures autonomes :** l'apprentissage profond est utilisé pour la détection d'objets, la segmentation et la reconnaissance de panneaux de signalisation dans les voitures autonomes.\n*   **Diagnostic médical :** l'apprentissage profond est utilisé pour analyser les images médicales afin de détecter des maladies telles que le cancer.\n*   **Surveillance :** l'apprentissage profond est utilisé pour la détection d'anomalies et la reconnaissance des activités dans les vidéos de surveillance.\n*   **Commerce électronique :** l'apprentissage profond est utilisé pour la reconnaissance de produits et la recommandation de produits dans les sites de commerce électronique.\n\nJ'espère que cela vous aidera ! N'hésitez pas à me poser d'autres questions."
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "multilingual",
    "test_name": "chinese_question",
    "prompt": "在机器翻译中最大的挑战是什么？",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 47.12865209579468,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Okay, let's break down the biggest challenges in machine translation. It’s a fascinating and complex field! Here's a breakdown of the key hurdles, categorized for clarity:\n\n**1. Linguistic Complexity – The Core Challenges:**\n\n* **Ambiguity:** This is *the* biggest problem. Words and phrases can have multiple meanings depending on context.  A single word in the source language can translate to several different words in the target language.  For example, the word \"bank\" can refer to a financial institution or the side of a river.  Machines struggle to consistently choose the correct interpretation.\n* **Idioms and Figurative Language:**  Idioms (“kick the bucket,” “raining cats and dogs”) and metaphors don’t translate literally.  A machine needs to recognize these and find an equivalent expression in the target language, which is incredibly difficult.\n* **Word Order Variation:** Languages have different grammatical structures. English is relatively fixed in its word order (Subject-Verb-Object), while languages like Japanese or Korean are much more flexible.  This requires the translation system to understand the *relationships* between words, not just their sequence.\n* **Morphology (Word Structure):**  Some languages (like German or Turkish) have very complex word formation rules – words can change dramatically depending on their grammatical function (case, tense, etc.). This makes it hard for a machine to parse and translate accurately.\n* **Syntax (Sentence Structure):**  Even beyond word order, the way sentences are constructed varies greatly.  A machine needs to understand the hierarchical structure of a sentence to translate it correctly.\n\n\n**2. Data-Related Challenges – The Fuel for the Machine:**\n\n* **Data Scarcity for Low-Resource Languages:**  Machine translation models are trained on massive amounts of parallel text (text in two languages that are translations of each other).  Many languages (e.g., many indigenous languages, less-spoken European languages) have very little available parallel data. This severely limits the quality of translation for these languages.\n* **Data Quality:**  Even with lots of data, the quality matters. If the parallel text is poorly translated, inaccurate, or contains errors, the machine translation system will learn those errors.\n* **Domain Specificity:** A model trained on general news articles might not perform well when translating legal documents or medical texts.  Specialized vocabulary and phrasing require domain-specific training data.\n\n**3. Context and World Knowledge – Going Beyond the Literal"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "multilingual",
    "test_name": "arabic_question",
    "prompt": "ما هي تطبيقات الذكاء الاصطناعي في مجال الرعاية الصحية؟",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 46.88947558403015,
    "gpu_after_load_mb": 8201.845703125,
    "output": "بالتأكيد، إليك بعض تطبيقات الذكاء الاصطناعي في مجال الرعاية الصحية:\n\n**1. التشخيص:**\n*   **تحليل الصور الطبية:** تستخدم خوارزميات الذكاء الاصطناعي لتحليل صور الأشعة السينية والتصوير بالرنين المغناطيسي والتصوير المقطعي المحوسب للكشف عن الأمراض مثل السرطان وأمراض القلب وأمراض الأوعية الدموية.\n*   **التشخيص المبكر:** يمكن للذكاء الاصطناعي تحليل البيانات السريرية لتحديد المرضى المعرضين لخطر الإصابة بمرض معين، مما يسمح بالتدخل المبكر.\n*   **تحليل علم الأمراض:** يمكن للذكاء الاصطناعي تحليل عينات الأنسجة لتحديد الخلايا السرطانية بدقة وسرعة.\n\n**2. العلاج:**\n*   **الطب الشخصي:** يمكن للذكاء الاصطناعي تحليل البيانات الجينية والبيانات السريرية للمريض لتحديد العلاج الأكثر فعالية له.\n*   **الروبوتات الجراحية:** تستخدم الروبوتات التي تعمل بالذكاء الاصطناعي لإجراء عمليات جراحية دقيقة مع تقليل التدخل الجراحي.\n*   **تطوير الأدوية:** يمكن للذكاء الاصطناعي تسريع عملية اكتشاف وتطوير الأدوية من خلال تحليل البيانات الكيميائية والبيولوجية.\n\n**3. إدارة الرعاية الصحية:**\n*   **مساعدات افتراضية للمرضى:** يمكن للمساعدات الافتراضية التي تعمل بالذكاء الاصطناعي الإجابة على أسئلة المرضى، وتذكيرهم بتناول الأدوية، وتوفير الدعم العاطفي.\n*   **إدارة السجلات الطبية:** يمكن للذكاء الاصطناعي تنظيم السجلات الطبية وتحليلها لتحديد الاتجاهات وتحسين جودة الرعاية.\n*   **تحسين كفاءة المستشفيات:** يمكن للذكاء الاصطناعي تحسين تخصيص الموارد وتقليل أوقات الانتظار وتحسين سير العمل في المستشفيات.\n\n**4. البحث والتطوير:**\n*   **تحليل البيانات الضخمة:** يمكن للذكاء الاصطناعي تحليل كميات هائلة من البيانات الطبية لتحديد الأنماط"
  },
  {
    "model_id": "google/gemma-3-4b-it",
    "category": "multilingual",
    "test_name": "german_question",
    "prompt": "Wie unterscheidet sich überwachtes Lernen vom unüberwachten Lernen?",
    "load_time_s": 9.74690842628479,
    "inference_time_s": 46.96996593475342,
    "gpu_after_load_mb": 8201.845703125,
    "output": "Gerne, hier ist ein Vergleich zwischen überwachtem und unüberwachtem Lernen:\n\n**Überwachtes Lernen**\n\n*   **Konzept:** Überwachtes Lernen ist wie das Lernen mit einem Lehrer. Ein Algorithmus wird mit einem Datensatz gefüttert, der aus Eingaben und den entsprechenden korrekten Ausgaben besteht. Das Ziel des Algorithmus ist es, eine Funktion zu lernen, die Eingaben auf Ausgaben abbildet.\n*   **Daten:** Benötigt gelabelte Daten, d. h. Daten, bei denen jede Eingabe mit der richtigen Ausgabe versehen ist.\n*   **Zweck:** Vorhersage oder Klassifizierung. Das Ziel ist es, eine Funktion zu lernen, die neue, ungesehene Eingaben korrekt vorhersagen kann.\n*   **Beispiele:**\n    *   Spam-Filter: Das Modell wird mit E-Mails gelabelt als \"Spam\" oder \"Nicht-Spam\" trainiert.\n    *   Bilderkennung: Das Modell wird mit Bildern gelabelt, die verschiedene Objekte darstellen (z. B. Katze, Hund, Auto).\n    *   Vorhersage von Hauspreisen: Das Modell wird mit historischen Daten über Häuser und deren Preise trainiert.\n\n**Unüberwachtes Lernen**\n\n*   **Konzept:** Unüberwachtes Lernen ist wie das Entdecken von Mustern in Daten ohne eine vorgegebene Ausgabe. Der Algorithmus wird mit einem Datensatz gefüttert, der nur Eingaben enthält, ohne dass die korrekten Ausgaben vorhanden sind.\n*   **Daten:** Benötigt ungelabelte Daten.\n*   **Zweck:** Entdeckung von Mustern und Strukturen in den Daten. Das Ziel ist es, die zugrunde liegende Struktur der Daten zu verstehen.\n*   **Beispiele:**\n    *   Kundensegmentierung: Das Modell gruppiert Kunden basierend auf ihrem Verhalten, ohne dass eine vordefinierte Segmentierung vorhanden ist.\n    *   Dimensionsreduktion: Das Modell reduziert die Anzahl der Variablen in einem Datensatz, während die wichtigsten Informationen erhalten bleiben.\n    *   Anomalieerkennung: Das Modell identifiziert ungewöhnliche Datenpunkte, die von der Norm abweichen.\n\nHier ist eine Tabelle, die die wichtigsten Unterschiede zusammenfasst:\n\n| Merkmal             | Überwachtes Lernen |"
  }
]