sileod's picture
Upload folder using huggingface_hub
0292f9d verified
|
Raw
History Blame Contribute Delete
2.41 kB
metadata
title: Reasoning Core Environment Server
emoji: 🧠
colorFrom: green
colorTo: blue
sdk: docker
pinned: false
app_port: 8000
base_path: /web
tags:
  - openenv
  - agent-environment
  - reasoning
  - reinforcement-learning
  - evaluation
  - symbolic-reasoning

Reasoning Core Environment

An OpenEnv environment for formally verifiable symbolic reasoning across logic, mathematics, planning, syntax, and related procedural domains.

Tasks come from reasoning-core/formal-reasoning-env and are scored by the task-specific evaluators in reasoning-core.

Use The Hosted Environment

from reasoning_core_env import ReasoningCoreAction, ReasoningCoreEnv

with ReasoningCoreEnv(
    base_url="https://reasoning-core-reasoning-core-openenv.hf.space"
) as env:
    result = env.reset(split="train", seed=42, size=1000)
    print(result.observation.prompt)

    result = env.step(ReasoningCoreAction(answer="<answer>...</answer>"))
    print(result.reward)

Each episode has one action:

  1. reset() returns a symbolic reasoning prompt.
  2. step(ReasoningCoreAction(answer=...)) scores the answer and ends the episode.

Plain answers and answers wrapped in <answer>...</answer> are accepted. Rewards are task-specific scores in the range 0 to 1.

The environment only serves pre-generated examples from the Hugging Face dataset. Rows whose task scorer is unavailable in the installed reasoning-core version are skipped, preventing deprecated or unsupported task types from reaching an episode.

Local Development

uv sync
uv run openenv validate
uv run openenv build -t reasoning-core-openenv

Run without Docker:

uv run server

The service exposes the interactive UI at /web, API documentation at /docs, health information at /health, and the persistent environment API at /ws.

Citation

If you use this environment, cite the Reasoning Core paper:

@article{reasoningcore2026,
  title={Reasoning Core: A Scalable Procedural Data Generation Suite for Symbolic Pre-training and Post-Training},
  author={Lacombe, Valentin and Quesnel, Valentin and Sileo, Damien},
  journal={arXiv preprint arXiv:2603.02208},
  year={2026},
  url={https://arxiv.org/abs/2603.02208}
}