🧮 NanoCalc-1M
NanoCalc-1M is a ultra-compact, character-level Seq2Seq Transformer based on the T5 architecture, specifically trained to perform arithmetic operations with high precision.
- Architecture: T5-based Encoder-Decoder
- Parameters: 0.99M
- Precision: Mixed Precision (BF16/FP16)
- Vocab: Character-level (0-9, +, -, *, /, =)
- Training Data: 2,000,000 synthetic samples (3-digit arithmetic)
- Max Input Length: 20 tokens
- Performance: ~97% Accuracy on 4-operation math (Validation Set)
Performance Chart
| Epoch | Training Loss | Val Accuracy | Status |
|---|
| 1 | 1.1420 | 54.89% | 🔴 Learnt Format |
| 2 | 0.3931 | 78.79% | 🟡 Learnt Digits |
| 5 | 0.1638 | 91.91% | 🟢 Learning subtleties |
| 9 | 0.1051 | 97.15% | 🔵 High Precision |
| 10 | 0.1004 | 97.73% | 🚀 Near Perfect |
How to use
To use this model, download model.pt and use.py and run it on any type of device with Python3.
Examples
Model loaded (Accuracy: 97.73% from epoch 10)
--- Mini Math Model interactive ---
Enter an arithmetic task (e.g. 15*15) or type 'exit' to quit this.
Task > 0*567
Model: 0 | Correct: 0 ✅
Task > 999+999
Model: 1998 | Correct: 1998 ✅
Task > 1/1
Model: 1 | Correct: 1 ✅
Task > 1684*8787
Model: 6398 | Correct: 14797308 ❌
Task > 124*598
Model: 2452 | Correct: 74152 ❌
Task > 12/68
Model: 4 | Correct: 0 ❌
Task > 123*123
Model: 499 | Correct: 15129 ❌
Task > 47*5
Model: 235 | Correct: 235 ✅
Task > 456+125
Model: 581 | Correct: 581 ✅
Task > 957-234
Model: 723 | Correct: 723 ✅
Task > 120-7650
Model: -550 | Correct: -7530 ❌
Task > 450-750
Model: -300 | Correct: -300 ✅
Task > 453-97
Model: 356 | Correct: 356 ✅
Task > 129-462
Model: -333 | Correct: -333 ✅
Task > 8*8
Model: 64 | Correct: 64 ✅
Task > 54*54
Model: 2916 | Correct: 2916 ✅
Task > 102*78
Model: 748 | Correct: 7956 ❌
Task > 74*9
Model: 666 | Correct: 666 ✅
Task > 103-34
Model: 69 | Correct: 69 ✅
Overall accuracy
The overall accuracy after 10 epochs of training is ~97% for tasks with max. 3 digits each like 74*9 or 103-34.
Limitations
The can't do:
- Tasks with more than 3 digits like
3984-125
- Multiplication tasks with numbers above 99 like
293*21
- Complex tasks
Training
We trained for 10 epochs (~20 minutes of training on Kaggle 2x T4) with 2 million randomly generated training samples.
Final thoughts
We may be releasing an improved version of this that can solve really complex tasks and much more...stay tuned!
