open-thought/reasoning-gym
1
0
Fork 0
mirror of https://github.com/open-thought/reasoning-gym.git synced 2026-04-29 17:35:16 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
mirror of github.com/open-thought/reasoning-gym
36 commits 21 branches 5 tags 37 MiB
Find a file
Andreas Koepf (aider) 432c9436f7 fix: Correct PatternRule.apply() method to properly handle sequence operations 
This commit message captures the essence of the change: fixing the implementation of the apply() method in the PatternRule class to correctly handle sequence operations and indexing.

The key changes are:
1. Use `sequence[position]` instead of `sequence[position - 1]`
2. Adjust PREV_PLUS condition to use `position > 0`
3. Use `sequence[position - 1]` for previous element reference

Would you like me to elaborate on the specific changes or rationale?
2025-01-23 13:52:47 +01:00
reasoning_gym fix: Correct PatternRule.apply() method to properly handle sequence operations 2025-01-23 13:52:47 +01:00
tests fix: Correct PatternRule.apply() method to properly handle sequence operations 2025-01-23 13:52:47 +01:00
.gitignore build: Initialize reasoning_gym package structure with packaging and development setup 2025-01-23 10:50:54 +01:00
.pre-commit-config.yaml feat: Add Black and isort pre-commit hooks with line length configuration 2025-01-23 11:02:13 +01:00
LICENSE Initial commit 2025-01-23 09:39:53 +01:00
pyproject.toml feat: Add Black and isort pre-commit hooks with line length configuration 2025-01-23 11:02:13 +01:00
python fix: Correct PatternRule.apply() method to properly handle sequence operations 2025-01-23 13:52:47 +01:00
README.md feat: Add arithmetic_dataset() factory function to basic_arithmetic.py 2025-01-23 12:47:01 +01:00
requirements-dev.txt feat: Add Black and isort pre-commit hooks with line length configuration 2025-01-23 11:02:13 +01:00

README.md

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.