| configs: | |
| - config_name: 1BT | |
| data_files: | |
| - split: train | |
| path: 1BT/train-*.parquet | |
| - split: test | |
| path: 1BT/test-*.parquet | |
| # 4-digit addition dataset (a + b = c, all <= 4 digits) | |
| Exhaustive enumeration of every (a, b) with a, b in [0, 9999] and | |
| a + b <= 9999 (so c is also <= 4 digits). 50,005,000 ordered pairs, | |
| commutative-safe 90/10 train/test split. | |
| Train ~45M examples (~570M tokens under Llama-3 tokenizer); test ~5M examples. | |