Summary

Aviary is a gymnasium for defining and running Reinforcement Learning (RL) environments for language agents, focusing on challenging scientific tasks. It offers pre-built environments for math, biology, and literature search, enabling agent benchmarking and training. Integrated with the LDP library, Aviary facilitates custom agent creation and evaluation.

How It Works

Aviary abstracts RL environments into a standardized interface with reset and step methods, managing state and tool interactions. Communication uses OpenAI-standard messages, supporting text and images. Agents, defined via the sister LDP library as Language Decision Processes, issue tool requests, and environments respond with observations, rewards, and status. This promotes modularity for complex agent development.

Quick Start & Requirements

• Install: pip install fhaviary
• With environments: pip install 'fhaviary[gsm8k,hotpotqa,labbench,lfrqa,notebook]'
• Prerequisites: Python environment; LLM API access (e.g., GPT-4o) for agent execution. Specific hardware (GPU/CUDA) may be needed for agent inference/training.
• Resources: Tutorials, documentation, and overview links are in the README.

Highlighted Details

• Diverse Environments: Pre-built for GSM8k (math), HotpotQA (knowledge), LAB-Bench (biology), LFRQA (literature search), and notebook execution.
• Flexible Definition: Supports class-based Environment subclassing and functional definition via decorators (@fenv.start, @my_env.tool).
• Advanced Tooling: Handles sophisticated tool signatures, multiline docstrings for LLM descriptions, and rich message formatting (text/images).
• Tool Inspection Server: A utility (aviary tools [env name]) allows viewing environment tools and their descriptions.

Maintenance & Community

Associated with a research paper and multiple contributors. Direct links to community channels (Discord, Slack), roadmaps, or sponsorship details are not provided in the README.

Licensing & Compatibility

The specific open-source license is not stated in the README, requiring further investigation for commercial use. It is designed to integrate with the LDP library.

Limitations & Caveats

Aviary's full utility requires integration with its sister library, LDP. The absence of a stated license is a critical adoption blocker. No mention of alpha/beta status, known bugs, or unsupported platforms.