Reasoning core ◉

reasoning-core is a text-based RLVR for LLM reasoning training. It is centered on expressive symbolic tasks, including full fledged FOL, formal mathematics with TPTP, formal planning with novel domains, and syntax tasks.

🤗 https://hf.co/datasets/reasoning-core/rc1

Prime Environment Hub

#!pip install uv #install uv if needed
!uv tool install prime --with openai  -q
!uv tool run prime -- env install sileod/reasoning-core-env

from verifiers import load_environment
import os; from openai import OpenAI

env = load_environment("reasoning-core-env")

client = OpenAI( base_url="https://openrouter.ai/api/v1", api_key=os.getenv("OPENROUTER_API_KEY")) #🔑
results = env.evaluate(client=client, model="gpt-4.1-mini", num_examples=20, rollouts_per_example=1)
df=env.make_dataset(results).to_pandas()

Standalone

pip install reasoning_core

from reasoning_core import list_tasks, get_task, score_answer

T = get_task('arithmetics')()
x = T.generate_example()
assert score_answer(x.answer, x)==1

Generation

Run bash run_generate.sh for multi-threaded generation to json files (readable by Huggingface Datasets).

Reasoning gym

We use a custom interface, leaner than reasoning-gym (RG). But our tasks, which are all orthogonal to RG, can be imported in it.

import reasoning_gym

from reasoning_core import register_to_reasoning_gym
register_to_reasoning_gym()

specs = [
    # here, leg_counting tasks will make up two thirds of tasks
    DatasetSpec(name='leg_counting', weight=2, config={}),  #from reasoning_gym 🏋
    DatasetSpec(name='arithmetics', weight=2, config={}),  #from reasoning_core ◉
]
D=reasoning_gym.create_dataset('composite', size=10, seed=42, datasets=specs)

Citation

@misc{reasoningcore2025,
      title={Reasoning Core: A Scalable RL Environment for LLM Symbolic Reasoning}, 
      author={Valentin Lacombe and Valentin Quesnel and Damien Sileo},
      year={2025},
      eprint={2509.18083},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2509.18083}, 
}