deqing/addition_dataset

---
configs:
- config_name: test
data_files:
- split: test
path: test/test-*.parquet
- config_name: 1BT
data_files:
- split: train
path: 1BT/train-*.parquet
- config_name: 10BT
data_files:
- split: train
path: 10BT/train-*.parquet
- config_name: 3MT-3digit
data_files:
- split: train
path: 3MT-3digit/train-*.parquet
- split: test
path: 3MT-3digit/test-*.parquet
---

# Addition Dataset

Addition problems in the format `{a} + {b} = {c}`.

## Subsets

- **test**: 5K held-out evaluation examples (operands >= 10, i.e. min 2 digits)
- **1BT**: ~85M training examples (~1 billion tokens under Llama-3 tokenizer)
- **10BT**: ~850M training examples (~10 billion tokens)
- **3MT-3digit**: Exhaustive single-token addition: all (a, b) with a, b in [0, 999] and a+b <= 999. 500,500 ordered pairs, ~3M tokens. All of a, b, c are single tokens. Symmetry-safe train/test split (10% test).

## Deduplication

- Commutative dedup: if `a + b = c` exists, `b + a = c` is excluded (1BT/10BT)
- Test exclusion: both orderings of test-set pairs are excluded from train splits
- 3MT-3digit: both orderings always in the same split (no commutative leakage)

## Usage

```python
from datasets import load_dataset

train = load_dataset("deqing/addition_dataset", "1BT", split="train")
test = load_dataset("deqing/addition_dataset", "test", split="test")

# Single-token exhaustive (0-999, a+b<=999)
train_3d = load_dataset("deqing/addition_dataset", "3MT-3digit", split="train")
test_3d = load_dataset("deqing/addition_dataset", "3MT-3digit", split="test")
```