Aviary ¹ is a gymnasium for defining custom language agent RL environments. The library features pre-existing environments on math ², general knowledge ³, biological sequences ⁴, scientific literature search ⁵, and protein stability. Aviary is designed to work in tandem with its sister library LDP (https://github.com/Future-House/ldp) which enables the user to define custom language agents as Language Decision Processes. See the following tutorial for an example of how to run an LDP language agent on an Aviary environment.

Overview | Getting Started | Documentation | Paper

• We have a new environment to run Jupyter notebooks at packages/notebook.

A pictorial overview of the five implemented Aviary environments and the language decision process framework.

To install aviary (note fh stands for FutureHouse):

pip install fhaviary

To install aviary together with the incumbent environments:

pip install 'fhaviary[gsm8k,hotpotqa,litqa,lfrqa,notebook]'

To run the tutorial notebooks:

pip install "fhaviary[dev]"

For local development, please see CONTRIBUTING.md.

1. Building a Custom Environment in Aviary
2. Building a GSM8K Environment in Aviary
3. Creating Language Agents to Interact with Aviary Environments
4. Evaluate a Llama Agent on GSM8K

The example below walks through defining a custom environment in Aviary. We define a simple environment where an agent takes actions to modify a counter. The example is also featured in the following notebook.

from collections import namedtuple
from aviary.core import Environment, Message, ToolRequestMessage, Tool

# State in this example is simply a counter
CounterEnvState = namedtuple("CounterEnvState", ["count"])


class CounterEnv(Environment[CounterEnvState]):
    """A simple environment that allows an agent to modify a counter."""

    async def reset(self):
        """Initialize the environment with a counter set to 0. Goal is to count to 10"""
        self.state = CounterEnvState(count=0)

        # Target count
        self.target = 10

        # Create tools allowing the agent to increment and decrement counter
        self.tools = [
            Tool.from_function(self.incr),
            Tool.from_function(self.decr),
        ]

        # Return an observation message with the counter and available tools
        return [Message(content=f"Count to 10. counter={self.state.count}")], self.tools

    async def step(self, action: ToolRequestMessage):
        """Executes the tool call requested by the agent."""
        obs = await self.exec_tool_calls(action)

        # The reward is the square of the current count
        reward = int(self.state.count == self.target)

        # Returns observations, reward, done, truncated
        return obs, reward, reward == 1, False

    def incr(self):
        """Increment the counter."""
        self.state.count += 1
        return f"counter={self.state.count}"

    def decr(self):
        """Decrement the counter."""
        self.state.count -= 1
        return f"counter={self.state.count}"

Following the definition of our custom environment, we can now evaluate a language agent on the environment using Aviary's sister library LDP (https://github.com/Future-House/ldp).

from ldp.agent import Agent
from ldp.graph import LLMCallOp
from ldp.alg import RolloutManager


class AgentState:
    """A container for maintaining agent state across interactions."""

    def __init__(self, messages, tools):
        self.messages = messages
        self.tools = tools


class SimpleAgent(Agent):
    def __init__(self, **kwargs):
        self._llm_call_op = LLMCallOp(**kwargs)

    async def init_state(self, tools):
        return AgentState([], tools)

    async def get_asv(self, agent_state, obs):
        """Take an action, observe new state, return value"""
        action = await self._llm_call_op(
            config={"name": "gpt-4o", "temperature": 0.1},
            msgs=agent_state.messages + obs,
            tools=agent_state.tools,
        )
        new_state = AgentState(
            messages=agent_state.messages + obs + [action],
            tools=agent_state.tools,
        )
        # Return action, state, value
        return action, new_state, 0.0


# Create a simple agent and perform rollouts on the environment

# Endpoint can be model identifier e.g. "claude-3-opus" depending on service
agent = SimpleAgent(config={"model": "my_llm_endpoint"})

runner = RolloutManager(agent=agent)

trajectories = await runner.sample_trajectories(
    environment_factory=CounterEnv,
    batch_size=2,
)

Below we expand on some of the core components of the Aviary library together with more advanced usage examples.

An environment should have two methods, env.reset and env.step:

obs_msgs, tools = await env.reset()
new_obs_msgs, reward, done, truncated = await env.step(action_msg)

Communication is achieved through messages.

The action_msg is an instance of ToolRequestMessage which comprises one or more calls to the tools returned by env.reset method.

The obs_msgs are either general obseravation messages or instances of ToolResponseMessage returned from the environment. while reward is a scalar value, and done and truncated are Boolean values.

We explain the message formalism in further detail below.

Communication between the agent and environment is achieved via messages. We follow the OpenAI standard. Messages have two attributes:

msg = Message(content="Hello, world!", role="assistant")

The content attribute can be a string but can also comprise objects such as images. For example, the create_message method can be used to create a message with images:

from PIL import Image
import numpy as np

img = Image.open("your_image.jpg")
img_array = np.array(img)

msg = Message.create_message(role="user", text="Hello, world!", images=[img_array])

In this case, content will be a list of dictionaries with the keys text and image_url.

{
    {"type": "text", "text": "Hello World!"},
    {"text": "image_url", "image_url": "data:image/png;base64,{base64_image}"},
}

The role, see the table below. You can change around roles as desired, except for tool which has a special meaning in aviary.

The Message class is extended in ToolRequestMessage and ToolResponseMessage to include the relevant tool name and arguments.

If you need more control over Environments and tools, you may wish to subclass Environment. We illustrate this with an example environment in which an agent is tasked to write a story.

We subclass Environment and define a state. The state consists of all variables that change per step that we wish to bundle together. It will be accessible in tools, so you can use state to store information you want to persist between steps and tool calls.

from pydantic import BaseModel
from aviary.core import Environment


class ExampleState(BaseModel):
    reward: float = 0
    done: bool = False


class ExampleEnv(Environment[ExampleState]):
    state: ExampleState

We do not have other variables aside from state for this environment, although we could also have variables like configuration, a name, tasks, etc. attached to it.

We will define a single tool that prints a story. Tools may optionally take a final argument state which is the environment state. This argument will not be exposed to the agent as a parameter but will be injected by the environment (if part of the function signature).

def print_story(story: str, state: ExampleState):
    """Print a story.

    Args:
        story: Story to print.
        state: Environment state (hidden from agent).
    """
    print(story)
    state.reward = 1
    state.done = True

The tool is built from the following parts of the function: its name, its argument's names, the arguments types, and the docstring. The docstring is parsed to obtain a description of the function and its arguments, so be sure to match the syntax carefully.

Environment episode completion is indicated by setting state.done = True. This example terminates immediately - other termination conditions are also possible.

It is also possible make the function async - the environment will account for that when the tool is called.

Aviary also supports more sophisticated signatures:

• Multiline docstrings
• Non-primitive type hints (e.g. type unions)
• Default values
• Exclusion of info below \f (see below)

If you have summary-level information that belongs in the docstring, but you don't want it to be part of the Tool.info.description, add a r prefix to the docstring and inject \f before the summary information to exclude. This convention was created by FastAPI (docs).

def print_story(story: str | bytes, state: ExampleState):
    r"""Print a story.

    Extra information that is part of the tool description.

    \f

    This sentence is excluded because it's an implementation detail.

    Args:
        story: Story to print, either as a string or bytes.
        state: Environment state.
    """
    print(story)
    state.reward = 1
    state.done = True

Next we define the reset function which initializes the tools and returns one or more initial observations as well as the tools. The reset function is async to allow for database interactions or HTTP requests.

from aviary.core import Message, Tool


async def reset(self):
    self.tools = [Tool.from_function(ExampleEnv.print_story)]
    start = Message(content="Write a 5 word story and call print")
    return [start], self.tools

Next we define the step function which takes an action and returns the next observation, reward, done, and whether the episode was truncated.

from aviary.core import Message


async def step(self, action: Message):
    msgs = await self.exec_tool_calls(action, state=self.state)
    return msgs, self.state.reward, self.state.done, False

You will probably often use this specific syntax for calling the tools - calling exec_tool_calls with the action.

Optionally, we can define a function to export a snapshot of the environment and its state for visualization or debugging purposes.

from aviary.core import Frame


def export_frame(self):
    return Frame(
        state={"done": self.state.done, "reward": self.state.reward},
        info={"tool_names": [t.info.name for t in self.tools]},
    )

If an environment can be instantiated without anything other than the task (i.e., it implements from_task), you can start a server to view its tools:

pip install fhaviary[server]
aviary tools [env name]

This will start a server that allows you to view the tools and call them, viewing the descriptions/types and output that an agent would see when using the tools.

Below we list some pre-existing environments implemented in Aviary:

Included with some environments are collections of problems that define training or evaluation datasets. We refer to these as TaskDatasets, e.g. for the HotpotQADataset subclass of TaskDataset:

from aviary.envs.hotpotqa import HotPotQADataset

dataset = HotPotQADataset(split="dev")

An alternative way to create an environment is using the functional interface, which uses functions and decorators to define environments. Let's define an environment that requires an agent to write a story about a particular topic by implementing its start function:

from aviary.core import fenv


@fenv.start()
def my_env(topic):
    # return the first observation and starting environment state
    # (empty in this case)
    return f"Write a story about {topic}", {}

The start decorator begins the definition of an environment.

The function, my_env, takes an arbitrary input and returns a tuple containing the first observation and any information you wish to store about the environment state (used to persist/share information between tools).

The state will always have an optional reward and a Boolean done that indicate if the environment episode is complete. Next we define some tools:

@my_env.tool()
def multiply(x: float, y: float) -> float:
    """Multiply two numbers."""
    return x * y


@my_env.tool()
def print_story(story: str | bytes, state) -> None:
    """Print a story to the user and complete episode."""
    print(story)
    state.reward = 1
    state.done = True

The tools will be converted into objects visible for LLMs using the type hints and the variable descriptions. Thus, the type hinting can be valuable for an agent that uses it correctly. The docstrings are also passed to the LLM and is the primary means (along with the function name) for communicating the intended tool usage.

You can access the state variable in tools, which will have any fields you passed in the return tuple of start(). For example, if you returned {'foo': 'bar'}, then you could access state.foo in the tools.

You may stop an environment or set a reward via the state variable as shown in the second print_story tool. If the reward is not set, it is treated as zero. Next we illustrate how to use our environment:

env = my_env(topic="foo")
obs, tools = await env.reset()

If Aviary is useful for your work please consider citing the following paper:

@article{Narayanan_Aviary_training_language_2024,
  title   = {{Aviary: training language agents on challenging scientific tasks}},
  author  = {
    Narayanan, Siddharth and Braza, James D. and Griffiths, Ryan-Rhys and
    Ponnapati, Manvitha and Bou, Albert and Laurent, Jon and Kabeli, Ori and
    Wellawatte, Geemi and Cox, Sam and Rodriques, Samuel G. and White, Andrew
    D.
  },
  year    = 2024,
  month   = dec,
  journal = {preprint},
  doi     = {10.48550/arXiv.2412.21154},
  url     = {https://arxiv.org/abs/2412.21154}
}