arithmark-2.0-train / README.md
TobiasLogic's picture
Add dataset card
4008aae verified
|
Raw
History Blame Contribute Delete
1.31 kB
metadata
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - arithmetic
  - arithmark
size_categories:
  - 1M<n<10M

arithmark-2.0-train

The training set behind arithmark-2.0-scratch-19m, a small character level model trained from scratch that scores 97.40% on ArithMark 2.0.

It holds 3 million procedurally generated integer arithmetic problems in the same format as ArithMark 2.0. Each expression ends in "=" and has one correct answer. Division is always exact and answers are non negative.

Every one of the 2500 expressions in the ArithMark 2.0 benchmark was removed from this set, so a model trained here is still evaluated on problems it has not seen.

Format

{"ctx": "3 * 9 + 12 / 1 =", "answer": "39", "text": "3 * 9 + 12 / 1 = 39"}
  • ctx: the expression, ending in "="
  • answer: the correct integer answer as a string
  • text: ctx and answer joined, ready for language model training

Distribution

Problems follow the ArithMark 2.0 mix. About half are single operator, the rest use two or three operators, and roughly a quarter use parentheses. Operators are addition, subtraction, multiplication and exact division. Operand ranges match the benchmark.

License

Apache 2.0.