Open-RL
Description
Open-RL is an environment sourced from the Turing Enterprises dataset for evaluating AI agents on self-contained, verifiable STEM reasoning problems. Problems span Physics, Mathematics, Chemistry, and Biology, requiring multi-step reasoning, symbolic manipulation, and numerical computation.
Capabilities
- Solving complex STEM problems requiring multi-step reasoning
- Symbolic manipulation and algebraic simplification
- Numerical computation and derivations
- Cross-domain scientific reasoning
License
Tasks
There are 40 tasks in the train split, covering:
- Physics: Astrophysics, electromagnetism, quantum mechanics, condensed matter, classical mechanics
- Mathematics: Number theory, special functions, combinatorics, analysis
- Chemistry: General, medicinal, inorganic chemistry
- Biology: Molecular biology, immunology, neurobiology, physiology, microbiology
Reward Structure
Binary reward (0 or 1) based on answer correctness. An LLM grader (gpt-5-mini) checks semantic/symbolic equivalence between the submitted answer and ground truth.
Data
Data is sourced from the TuringEnterprises/Open-RL dataset on Hugging Face.
Tools
Single tool:
answer(answer: str)- Submit your solution to be graded
Time Horizon
Open-RL is a single-turn environment. Each task requires exactly one tool call to submit an answer. The agent receives a problem, performs reasoning, and submits its final answer.
Other Environment Requirements
- OpenAI API Key: Required for LLM-based grading
Safety
Open-RL presents minimal safety risks. Agents interact only with static STEM problems and submit text answers for grading. There is no network access, filesystem interaction, or execution of agent-generated code. The environment does not involve real-world actions, external systems, or other agents.
Citations
@dataset{OpenRL2024,
author = {Turing Enterprises},
title = {Open-RL: Self-contained, verifiable STEM reasoning problems},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/TuringEnterprises/Open-RL}
}