Reasoning Gym

We are building a python library of procedural dataset generators and algorithmically verifiable reasoning environments for training Reasoning Models with reinforcement learning (RL).

The goal is to generate virtually infinite data with adjustable complexity.

Quick Start

from reasoning_gym.arithmetic import ChainSum, ChainSumConfig

# configure a simple arithmetic task generator
config = ChainSumConfig(
    min_terms=2,
    max_terms=6,
    min_digits=1,
    max_digits=4,
    allow_negation=False, # Only positive numbers
    size=5,               # virtual size of dataset
    seed=42               # deterministic results
)

# create the dataset
dataset = ChainSum(config)

# print some examples
for item in dataset:
    print(item)

Example output:

{'question': '4 + 3 =', 'answer': '7', 'metadata': {'num_terms': 2, 'num_digits': 1, 'expression': '4 + 3'}}
{'question': '812 + 880 =', 'answer': '1692', 'metadata': {'num_terms': 2, 'num_digits': 3, 'expression': '812 + 880'}}
{'question': '2 + 6 + 3 + 4 + 0 =', 'answer': '15', 'metadata': {'num_terms': 5, 'num_digits': 1, 'expression': '2 + 6 + 3 + 4 + 0'}}
{'question': '8995 - 5221 + 2341 + 5967 =', 'answer': '12082', 'metadata': {'num_terms': 4, 'num_digits': 4, 'expression': '8995 - 5221 + 2341 + 5967'}}
{'question': '1654 + 4744 =', 'answer': '6398', 'metadata': {'num_terms': 2, 'num_digits': 4, 'expression': '1654 + 4744'}}

Generator / Environment Ideas

• math tasks
• algorithmic tasks (counting, sorting, re-ordering, ..)
• logic riddles
• logic inductive programming tasks
• ARC-AGI synthetic riddles

Call for Contributions

If you have ideas for additional procedural dataset generators or please create an issue here.

Or contact us in the #arc-agi-2 channel of the GPU-Mode discord server.