Math Expert

Arithmetic, algebra, and mathematical reasoning

Purpose

The Math Expert validates mathematical claims, solves arithmetic problems, and performs algebraic reasoning within the BeastBullet system.

Key Function: Provide structured mathematical validation (VALID/INVALID/UNCERTAIN) for reasoning pipelines.

Architecture

Input: 256-dim embeddings (from shared encoder)
Hidden: 256-dim, 3 layers
Output: 3 classes (VALID, INVALID, UNCERTAIN)
Parameters: 1.3M
Size: 5.2MB

Model Type: Feed-forward neural network with ReLU activation

Training

Dataset

Primary: Custom math dataset (1000+ examples)
Teacher Model: Ollama/Mistral
Categories: Arithmetic, algebra, percentages, fractions, word problems

Methodology

Training Samples: 1000+
Validation Samples: 200+
Epochs: 10
Batch Size: 32
Learning Rate: 1e-4
Accuracy: 85.3% (validation)

Performance

Validation Metrics

Overall Accuracy: 85.3%
Arithmetic: 92% accuracy
Algebra: 81% accuracy
Word Problems: 78% accuracy

Inference

Latency: <10ms (CPU)
Deterministic: 100% (same input → same output)

Output Format

{
  "class": "VALID",  # or "INVALID", "UNCERTAIN"
  "confidence": 0.89,
  "logits": [0.05, 0.06, 0.89]
}

Use Cases

✅ Good Use Cases

Validating mathematical claims
Solving basic arithmetic (addition, subtraction, multiplication, division)
Percentage calculations
Simple algebra

❌ Not Suitable For

Advanced calculus or differential equations
Symbolic mathematics
Theorem proving
Multi-step complex proofs

Integration Example

from beastbullet.experts import MathExpert
from beastbullet.core import SharedEncoder

encoder = SharedEncoder("shared_encoder_v1.0.pt")
math_expert = MathExpert("math_expert_v1.0.pt")

# Validate calculation
claim = "15% of 240 is 36"
embedding = encoder.encode(claim)
result = math_expert.validate(embedding)

print(result)
# {"class": "VALID", "confidence": 0.92}

SetMD
/

beastbullet-experts