stream-dataset 0.1.0

STREAM (Sequential Tasks Review to Evaluate Artificial Memory) is a dataset of 12 diverse sequential tasks to assess neural networks’ memory. Scalable in complexity and sequence length, it covers pattern completion, copy tasks, forecasting, bracket matching, and sorting—ideal for comparing architectures on memory retention and sequential reasoning.

Stream Dataset

A comprehensive dataset suite for evaluating sequence modeling capabilities of neural networks, particularly focusing on memory, long-term dependencies, and temporal reasoning tasks.

🚀 Features

• 12 Diverse Tasks: From simple memory tests to complex pattern recognition
• Multiple Difficulty Levels: Small, medium and large configurations. Advice : if you're building an architecture, you should begin with small.
• Unified Interface: Consistent API across all tasks with standardized evaluation metrics
• Ready-to-Use: Pre-configured datasets with train/validation/test splits
• Flexible: Support for both classification and regression tasks

📦 Installation

pip install stream-dataset

Or install from source:

git clone https://github.com/Naowak/stream-dataset.git
cd stream-dataset
pip install -e .

🎯 Quick Start

import stream_dataset as sd

# Build a task
task_data = sd.build_task('simple_copy', difficulty='small', seed=0)

# Access the data
X_train = task_data['X_train']  # Training inputs
Y_train = task_data['Y_train']  # Training targets
T_train = task_data['T_train']  # Prediction timesteps

# Train your model (example with dummy predictions)
Y_pred = your_model.predict(X_train)

# Evaluate performance
score = sd.compute_score(
    Y=Y_train, 
    Y_hat=Y_pred, 
    prediction_timesteps=T_train,
    classification=task_data['classification']
)
print(f"Score: {score}")

📚 Available Tasks

Memory Tests

Postcasting Tasks

• discrete_postcasting: Copy discrete sequences after a delay

  

• continuous_postcasting: Copy continuous sequences after a delay

  

Signal Processing

• sinus_forecasting: Predict frequency-modulated sinusoidal signals

  

• chaotic_forecasting: Forecast Lorenz system dynamics

  

Long-term Dependencies

• discrete_pattern_completion: Complete masked repetitive patterns (discrete)

  

• continuous_pattern_completion: Complete masked repetitive patterns (continuous)

  

• simple_copy: Memorize and reproduce sequences after delay + trigger

  

• selective_copy: Memorize only marked elements and reproduce them

  

Information Manipulation

• adding_problem: Add numbers at marked positions

  

• sorting_problem: Sort sequences according to given positions

  

• bracket_matching: Validate parentheses sequences

  

• sequential_mnist: Classify MNIST digits read column by column

  

🔧 Task Configuration

Each task supports three difficulty levels (the following numbers may vary in function of the task):

Small

• Reduced sequence lengths and sample counts
• Suitable for quick experiments and debugging
• Example: 100 training samples, sequences of ~50-100 timesteps

Medium

• Realistic problem sizes
• Suitable for thorough model evaluation
• Example: 1,000 training samples, sequences of ~100-200 timesteps

Large

• Big Data configurations
• Suitable for high-performance models
• Example: 10,000 training samples, sequences of ~200-500 timesteps

For the tasks that allow it, the difficulty is also adjusted by the dimensions of input & output.

# Small configuration (fast)
task_small = sd.build_task('bracket_matching', difficulty='small', seed=0)

# Medium configuration (thorough)
task_medium = sd.build_task('bracket_matching', difficulty='medium', seed=0)

# Large configuration (comprehensive)
task_large = sd.build_task('bracket_matching', difficulty='large', seed=0)

📊 Data Format

All tasks return a standardized dictionary:

{
    'X_train': np.ndarray,      # Training inputs [batch, time, features]
    'Y_train': np.ndarray,      # Training targets [batch, time, outputs]
    'T_train': np.ndarray,      # Training prediction timesteps [batch, n_predictions]
    'X_valid': np.ndarray,      # Validation inputs
    'Y_valid': np.ndarray,      # Validation targets  
    'T_valid': np.ndarray,      # Validation prediction timesteps
    'X_test': np.ndarray,       # Test inputs
    'Y_test': np.ndarray,       # Test targets
    'T_test': np.ndarray,       # Test prediction timesteps
    'classification': bool      # True for classification, False for regression : used to select the correspongind loss
}

🎨 Example: Complete Evaluation Pipeline

import stream_dataset as sd
import numpy as np
from MyModel import MyModel

def evaluate_model_on_all_tasks(model, difficulty='small'):
    """Evaluate a model on all available tasks."""
    
    results = {}
    task_names = [
        'simple_copy', 'selective_copy', 'adding_problem',
        'discrete_postcasting', 'continuous_postcasting',
        'discrete_pattern_completion', 'continuous_pattern_completion',
        'bracket_matching', 'sorting_problem', 'sequential_mnist',
        'sinus_forecasting', 'chaotic_forecasting'
    ]
    
    for task_name in task_names:
        print(f"Evaluating on {task_name}...")
        
        # Load task
        task_data = sd.build_task(task_name, difficulty=difficulty)
        
        # Train model (simplified)
        model = MyModel(...)
        model.train(task_data['X_train'], task_data['Y_train'], task_data['T_train'])
        # If you want your model to learn all timesteps, including the ones that are not evaluated :
        # Comment the previous line and uncomment the following one
        # model.train(task_data['X_train'], task_data['Y_train'])

        # Predict on test set
        Y_pred = model.predict(task_data['X_test'])
        
        # Compute score
        score = sd.compute_score(
            Y=task_data['Y_test'],
            Y_hat=Y_pred,
            prediction_timesteps=task_data['T_test'],
            classification=task_data['classification']
        )
        
        results[task_name] = score
        print(f"  Score: {score:.4f}")
    
    return results

# Usage
# results = evaluate_model_on_all_tasks(your_model, difficulty='medium')

🧪 Task Details

Memory and Copy Tasks

Test the model's ability to store and retrieve information over time.

Pattern Recognition Tasks

Evaluate pattern detection and completion capabilities with both discrete and continuous sequences.

Algorithmic Reasoning Tasks

Assess the model's ability to perform computations on stored information (adding, sorting).

Real-world Sequence Tasks

Include practical applications like MNIST digit classification and signal forecasting.

📈 Evaluation Metrics

• Classification tasks: Error rate (1 - accuracy)
• Regression tasks: Mean Squared Error (MSE)

Lower scores indicate better performance for both metrics.

🤝 Contributing

We welcome contributions! Please see our Contributing Guidelines for details.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

📚 Citation

If you use Stream Dataset in your research, please cite:

(paper in progress...)

@software{stream_dataset,
  title={Stream Dataset: Sequential Task Review to Evaluate Artificial Memory},
  author={Yannis Bendi-Ouis, Xavier Hinaut},
  year={2025},
  url={https://github.com/Naowak/stream-dataset}
}

🙏 Acknowledgments

• Inspired by classic sequence modeling datasets
• Built with NumPy and Hugging Face Datasets
• MNIST task uses the official MNIST dataset

📞 Support

• 🐛 Issues: GitHub Issues
• 💬 Discussions: GitHub Discussions

---

Stream Dataset - Advancing sequence modeling evaluation, one task at a time.