TrueMath
A 1-layer TrueACT model trained to generate chain-of-thought reasoning for arithmetic expressions with +, -, *, and parentheses.
Trained from scratch on purely random binary-tree expressions. Every non-root subexpression is parenthesized, forcing the model to learn step-by-step reduction of nested arithmetic.
Performance
| Test | Accuracy |
|---|---|
| Fixed 12-case benchmark | 12/12 (100%) |
| Random 500 expressions | 91.6% |
Errors are exclusively multi-digit arithmetic mistakes (e.g., 42×88=3524). The model's structural reasoning (parenthesis resolution, operator precedence, chain-of-thought decomposition) is near-perfect.
Files
| File | Description |
|---|---|
model.pt |
TorchScript-traced model for inference (no Python source needed) |
checkpoint.pt |
Original PyTorch checkpoint (requires the TrueACT architecture code to load) |
infer.py |
Standalone inference script |
plot.py |
Generate the training curves figure from the training log |
Usage
# Single prompt
python infer.py model.pt '((5*5)+(10*2))='
# Interactive mode
python infer.py model.pt
Prompt > (1+2)*3=
The model expects a prompt ending with = and generates the chain-of-thought steps:
Input: ((5*5)+(10*2))=
Output: ((5*5)+(10*2))=(25+(10*2))=(25+20)=45
Requirements
- Python 3.10+
- PyTorch 2.0+
No other dependencies.