deqing's picture
Fix YAML: 1BT config now has train + test splits (exhaustive 4-digit)
9934a51 verified
|
Raw
History Blame Contribute Delete
455 Bytes
metadata
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.