reasoning-core OpenReward environment

This folder contains an OpenReward-compatible ORS environment for reasoning-core.

Files

• server.py: OpenReward environment server implementation.
• requirements.txt: Python dependencies for local/dev builds.
• Dockerfile: Container spec used by OpenReward deployment.

Local run

cd reasoning_core/openreward/reasoning_core_env
pip install -r requirements.txt
python server.py

Quick local smoke test

from openreward import OpenReward

or_client = OpenReward()
env = or_client.environments.get(name="reasoningcore", base_url="http://localhost:8080")

print(env.list_splits())
print(env.list_tools())
print(env.list_tasks("train")[:1])

Configuration

Set optional environment variables before launch:

• RC_NUM_TRAIN (default 500)
• RC_NUM_TEST (default 50)
• RC_SEED (default 0)
• RC_PASS_THRESHOLD (default 0.98)
• RC_HF_DATASET (default reasoning-core/procedural-pile)
• RC_HF_REVISION (optional tag or commit of the dataset, to pin the tasks)
• RC_HF_CONFIG (optional dataset config name)

The task order is deterministic for fixed values of these variables.

Notes

• The environment exposes a single answer tool.
• The tool returns a human-readable result with a rounded reward (reward=0.000 format).
• "Accepted" defaults to reward >= 0.98 (configurable via RC_PASS_THRESHOLD).
• The tool accepts either plain-text answers or XML-wrapped answers (<answer>...</answer>), matching common evaluator output formats.
• By default, tasks are the first rows of reasoning-core/procedural-pile (pre-shuffled, all task families within the first few hundred rows). Rows whose task has no scorer in the installed reasoning-core are skipped. If the dataset has no native test split, test tasks are sampled from train.
• If Hugging Face loading fails, the environment falls back to deterministic procedural task generation.