Von

Von: An Open-Source Non-Autoregressive Decision Model
Von is an advanced System One decision model designed for high-speed decision-making in computational tasks. With capabilities to deliver responses in under 25 milliseconds, it offers a powerful, highly efficient alternative to traditional autoregressive models, addressing significant latency issues and memory overhead often associated with large language models (LLMs).

Overview

Unlike autoregressive models that process data token by token, Von performs classification, intent routing, and verification through direct inference. By leveraging a reflexive and parallel processing approach, Von significantly reduces execution time and improves accuracy in various application domains.

Key Capabilities

• Non-Autoregressive Parallelism: Evaluate multiple independent questions in a single forward pass, optimizing processing speed.
• SOTA Empirical Accuracy: Achieve a remarkable 91.23% accuracy on adversarial multi-hop reasoning benchmarks, surpassing many commercial alternatives.
• Calibrated Uncertainty: Output probabilities are attuned to reflect genuine predictive confidence through advanced calibration techniques.
• Hardware Agnostic Acceleration: Optimized for various hardware environments, including NVIDIA CUDA, AMD ROCm, and Apple Silicon MPS, ensuring broad accessibility.
• Protocol Parity: Fully compatible with the TypeSafe /v1/systemone specification, enhancing integration capabilities across different systems.

Empirical Benchmark

In independent evaluations, Von has demonstrated superior performance across a range of tasks, particularly in real-time scenarios.

Decision Primitives

Von formalizes decision-making through three key primitives:

1. Choice: Categorical decision analysis, computing normalized probability distributions.
2. Noul: Binary probability verification to estimate posterior probabilities effectively.
3. Score: Evaluating ordinal ratings through expected values across severity or quality levels.

Training Methodology & Calibration

Von utilizes an innovative training approach combining Reinforcement Learning with Calibration Distribution (RLCD) to enhance both classification precision and probabilistic accuracy, featuring meticulously crafted training datasets that ensure balanced learning across diverse contexts.

Use Cases

Von can be applied in multiple scenarios, such as IT support ticket classification, real-time gaming control via ViZDoom, security incident management, and operational triage, demonstrating versatility across various domains.

Example Usage

In Python, Von can be effortlessly implemented to conduct discrete decisions based on system states:

import von

result = von.decide(
    state="Database replication lag on cluster us-west-2 exceeded 45 seconds.",
    choices={
        "infrastructure": "Database, hardware, network, or server failures",
        "billing": "Invoices, payments, refunds, subscription queries",
        "feature_request": "Requests for new platform capabilities",
    },
)
print(result.choice)         # 'infrastructure'
print(result.confidence)     # 0.8412
print(result.probabilities)  # {'infrastructure': 0.9021, 'billing': 0.0489, ...}

Theoretical Homage

The name 'Von' acknowledges the legacies of John von Neumann and Ludwig von Mises, who contributed significantly to computing and decision theory.

For more information, visit the GitHub repository and explore the potential of the Von decision model.