MultivexAI/NumAdd-v1.0

Model Description

NumAdd-v1.0 is a lightweight feed-forward neural network (FNN) implemented in PyTorch for numerical sum prediction. Architecture: 2-input, 1-output, with two hidden layers (32, 64 neurons) and ReLU activations. Parameters: 2,273 trainable.

Evaluation

Benchmarked on 120,000 samples across five input magnitude ranges. Metrics: MAE, MSE, RMSE, R2.

Range (Input Max)	MAE	MSE	RMSE	R2
0-50	0.003	0.000	0.004	1.000
51-500	0.003	0.000	0.004	1.000
501-5000	0.004	0.000	0.006	1.000
5001-50000	0.016	0.003	0.050	1.000
50001-500000	0.1525	0.2377	0.4876	1.000
500001-50000000	12.947	2143.782	46.301	1.000

Limitations

Performance degrades significantly for large magnitude inputs (>50,000), evidenced by increased MAE/MSE, despite maintaining high R2.

Crucial Safety Warning, Disclaimers, & Liability Limitation

This model is an experimental statistical prototype provided strictly "as is" and "with all faults." The developers, distributors, and contributors disclaim all warranties, express or implied, including but not limited to any implied warranties of merchantability, fitness for a particular purpose, or non-infringement.

• Catastrophic Silent Failures (No Runtime Alerts): Unlike traditional deterministic algorithms, this neural network does not throw overflow, underflow, or computation errors when failing. When presented with values outside its optimal range, it will silently output plausible-looking but mathematically incorrect numbers. This makes it highly dangerous for automated systems, as failures cannot be caught using standard software exception handling.
• Total Absence of Formal Verification: No formal mathematical verification, static analysis, or correctness proofs have been performed on this neural network's weights. Its behavior is non-deterministic across different hardware backends (CPU vs. GPU) due to floating-point truncation variations and cannot be relied upon for any application requiring consistent or auditable mathematical outputs.
• Severe Extrapolation & Out-of-Distribution Degradation: As documented in the evaluation data, the model exhibits catastrophic degradation on inputs exceeding 50,000, with absolute errors climbing into double digits. It must be assumed to fail completely when presented with extremely large, extremely small, fractional, negative, or structurally complex numerical inputs.
• Statistical Metric Deception ($R^2$): Users are explicitly warned that the reported $R^2$ score of 1.000 is a statistical artifact of scale variance. It is mathematically deceptive and must not be used to justify the model’s safety, precision, or operational readiness.
• Arbitrary Code Execution (Security Vulnerability): The serialized model file is saved in PyTorch's default pickle-based format. Loading untrusted .pt or .pth files can execute arbitrary, malicious code on your system. To mitigate the risk of remote code execution (RCE) and system compromise, you must enforce strict sandboxing and use weights_only=True:

  torch.load("model.pth", map_location=device, weights_only=True)
  

• Complete Assumption of Risk: The end-user assumes 100% of the risk associated with integrating this model into any system. The creators shall not be held liable for any direct, indirect, incidental, special, or consequential damages (including but not limited to financial loss, hardware damage, or system downtime) arising from its use.