tessera-core

Tessera-Core: An Open Protocol for AI-to-AI Knowledge Transfer
Tessera-Core is an innovative protocol designed for transferring knowledge between advanced AI models, enhancing the ability of trained neural networks to share their learned experiences with untrained models across diverse architectures. This project introduces a unique framework that encodes knowledge as activation-level representations and facilitates a seamless transfer via a Universal Hub Space, ensuring scalability and fidelity through established metrics.

Key Features

• Cross-Architecture Knowledge Transfer: Enables communication and learning among models that follow different architectural designs, allowing them to share their insights without being architecture-dependent.
• Activation Patterns Transfer: Focuses on transferring behavioral patterns rather than raw weights, extracting essential operational features while maintaining architectural flexibility.
• Efficient Knowledge Encoding: Utilizes innovative algorithms to collect activation statistics, train shared encoders/decoders, and verify knowledge fidelity to ensure high transfer quality.

Quick Start Example

To initiate a knowledge transfer between a trained model and an untrained target, consider the following code snippet:

from tessera import ModeATransfer, TBFSerializer, QuantType

# Initialize transfer process
transfer = ModeATransfer(
    transmitter=trained_model,
    receiver=untrained_model,
    transmitter_id="model_a",
    receiver_id="model_b",
)

# Execute the transfer
token = transfer.execute(
    train_dataloader=train_loader,
    val_dataloader=val_loader,
    privacy_epsilon=1.0,
)

# Save the transfer token as a compact binary (TBF v1.1)
TBFSerializer.save("transfer.tbf", token, quant=QuantType.FLOAT16)

Benchmarking Results

Tessera has undergone rigorous testing across various architecture families— Transformer, MLP, Conv1D, and bidirectional LSTM — successfully validating knowledge transfers in different configurations. Key results include:

Understanding the Universal Hub Space

The Universal Hub Space (UHS) is critical in the knowledge transfer process, accommodating various dimensions depending on model complexity. Defaulting to 2048 dimensions, the UHS is adjustable for specific architectures to prevent information bottlenecks and ensure effective transfers.

Resources for Developers

For further understanding and contributions, developers can explore additional documentation including:

• Tessera Specification
• Privacy and Security Audit
• Integration with Model Context Protocol and Agent-to-Agent protocol

Tessera-Core stands as an essential tool for advancing AI capabilities through effective inter-model knowledge sharing. For more information, please visit the project repository.

Explore the potential of AI collaboration with Tessera-Core, bridging the gap across different neural network architectures to achieve enhanced learning outcomes.