Single Step environments are the most widespread type of environment when doing RL with an LLM as policy.

It’s called single step environment, because the agent only does one step. It gets a question sampled from the dataset (the initial state / observation) and then answers. The answer is then scored according to the reward function. Recently, rules-based reward functions, i.e. functions without any learnable parameters, have been successfully used to do RL with LLMs as as policy.

First, we have to load a dataset from which we will sample questions. The dataset can be either a StaticDataset, which is finite and the length is known at runtime, or it can be a BaseGenerator, which is an infinite supply of question - answer pairs, synthetically generated in some way (depending on the implementation).

For the sake of simplicity, we will start by loading the MATH dataset, remove unnecessary columns and rename the remaining ones, such that we can easily turn it into a StaticDataset, which SingleStepEnv can deal with.

First, install the CAMEL package with all its dependencies:

%pip install camel-ai[all]==0.2.46
from datasets import load_dataset

from camel.datasets import StaticDataset
from camel.logger import get_logger

logger = get_logger(__name__)


dataset = load_dataset("EleutherAI/hendrycks_math", "algebra")

# Preprocess
dataset["train"] = dataset["train"].rename_column("problem", "question")
dataset["train"] = dataset["train"].rename_column("solution", "final_answer")
dataset["train"] = dataset["train"].remove_columns(["type", "level"])

# This should now print "['question', 'final_answer']"
print(dataset["train"].column_names)
seed_dataset = StaticDataset(dataset['train'])

print("Example datapoint:", seed_dataset[0])

Next, we will have to define an extractor. An extractor takes the LLM response and extracts the verifiable part out of it. Extractors can be initialized with different strategies which modifies the extraction behavior.

In the case of the MATH dataset, the final answer is wrapped inside a \boxed{...}, hence we should use the pre-built BoxedStrategy.

Sadly, MATH answers are rather complicated and a more general Math verifier to compare, for example, equations has not yet been implemented. Hence, we shall prune the dataset to only contain those rows where the content of \boxed{...} is an int. For the sake of simplicity, we shall also prune the ground truthes to the direct answer (such that they are python expressions). That way, we can do simple verification using the vanilla PythonVerifier!

from camel.extractors import BaseExtractor, BoxedStrategy

# Initialize extractor
extractor = BaseExtractor([[BoxedStrategy()]])
await extractor.setup()

valid_datapoints = []

# Iterate through dataset, checking for datapoints with integer answers
for datapoint in seed_dataset:
    extracted_value = await extractor.extract(response=datapoint.final_answer)

    if not extracted_value:
        continue
    if extracted_value.isdigit() or (
        extracted_value.startswith('-') and extracted_value[1:].isdigit()
    ):
        valid_datapoints.append(
            {
                "question": datapoint.question,
                "final_answer": extracted_value,
            }
        )

# We should now have `1228` valid data points.
print(f"Number of datapoints with integer answers: {len(valid_datapoints)}")

filtered_dataset = StaticDataset(valid_datapoints, seed=42)

Let’s create a Python verifier to later compare answers. Since we are reusing the same extractor from before, the PythonVerifier will expect solutions to be wrapped in \boxed{...}, too.

from camel.verifiers import PythonVerifier

verifier = PythonVerifier(extractor=extractor)
await verifier.setup(uv=True)

Let’s now initialize the single step environment with our filtered dataset and our verifier. The verifier will later help with the correctness reward

We can then call env.reset() to draw from the initial state distribution and return an observation, which can then be fed into the agent.

from camel.environments import Action, SingleStepEnv

env = SingleStepEnv(filtered_dataset, verifier)

obs = await env.reset(seed=42)

print(obs)

The agent would then process this observation and select an action, which it would feed into the step function. An action in this case would simply be the answer to the question, wrapped in \boxed{} (since we initialized our verifier with an extractor that extracts from \boxed{...})

await env.step(Action(index=0, llm_response="\\boxed{5}"))

As you can see, the output of the step function contains the next observation (which in this case is just a placeholder, since the episode is over), a reward, as well as a reward dict, showing exactly which rubric brought which reward, a done flag, indicating that the episode is over and some additional info.

In this case, we get a reward of 1010, which is the reward for a correct final answer. This can be accessed and changed via the self.ACCURACY_REWARD attribute.

Since we did not implement any other reward components, such as a formatting reward, the accuracy reward is our total reward.