countdown-dataset / README.md
alexander-lu's picture
Update dataset card
b617441 verified
|
Raw History Blame Contribute Delete
1.8 kB
---
pretty_name: ES Heterogeneity Countdown
task_categories:
- text-generation
language:
- en
size_categories:
- 1M<n<10M
configs:
- config_name: default
data_files:
- split: train
path: train.jsonl
- split: test
path: test.jsonl
---
# ES Heterogeneity Countdown
Countdown arithmetic data used for Evolution Strategies experiments under
heterogeneous data allocation.
## Dataset splits
- `train`: approximately 3.79 million synthetically generated, solvable, and
deduplicated Countdown problems.
- `test`: 2,000 held-out Countdown problems from the original evaluation set.
Each example contains:
- `id`: example identifier
- `numbers`: input numbers that must each be used exactly once
- `target`: desired arithmetic result
- `context`: complete model prompt
- `solution`: generated witness expression, when available
## Task
Construct an expression using every supplied number exactly once. The permitted
operations are addition, subtraction, multiplication, and division. The
expression must equal the specified target.
The prompt requests reasoning in `<think>...</think>` tags and a final expression
in `<answer>...</answer>` tags.
## Data generation
Synthetic training instances are created by constructing a random valid
arithmetic expression and evaluating it to obtain a reachable target. Duplicate
`(sorted numbers, target)` puzzles are removed. Number sets present in the held-out
evaluation data are excluded from synthetic training.
The training pool contains both three-number and four-number problems and is
weighted toward the harder four-number setting.
## Evaluation
An answer is correct when it:
1. Uses every supplied number exactly once.
2. Uses only `+`, `-`, `*`, `/`, and parentheses.
3. Evaluates to the requested target.