Single step Environment is not deterministic for BaseGenerator Dataset

[Tanuj-Taneja1]

The reset method in SingleStepEnv (in single_step.py) does not actually use the provided seed argument when the BaseGenerator dataset is used.

Currently, each call to reset invokes the dataset’s async_sample method, which gives next datapoint every time — even if the same seed is passed. This means the environment does not reset deterministically as expected.
This should be relatively easy fix.

Also can we rename the BaseGenerator dataset to a more descriptive something like SyntheticDataset or DynamicDataset.

[1sarthakbhardwaj]

Hi @Wendong-Fan can you help here

[Tanuj-Taneja1]

to reproduce this issue you can just run the code below, also i am using camel version 0.2.75

import asyncio
from camel.datasets import BaseGenerator, DataPoint
from camel.environments import SingleStepEnv
from camel.verifiers import PythonVerifier


class SimpleGenerator(BaseGenerator):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)

    async def generate_new(self, n: int, **kwargs) -> None:
        for i in range(n):
            datapoint = DataPoint(
                question=f"What is {i} + {i}?",
                final_answer=str(i * 2),
            )
            self._data.append(datapoint)


async def main():
    generator = SimpleGenerator(buffer=6, seed=1) 
    verifier = PythonVerifier()  
    env = SingleStepEnv(dataset=generator, verifier=verifier)
    await env.setup()

    for i in range(5):
        obs = await env.reset(batch_size=2, seed=5)
        print(obs)
    await env.close()


if __name__ == "__main__":
    asyncio.run(main())

[Observation(question='What is 0 + 0?', context={}, metadata={}), Observation(question='What is 1 + 1?', context={}, metadata={})]
[Observation(question='What is 2 + 2?', context={}, metadata={}), Observation(question='What is 3 + 3?', context={}, metadata={})]
[Observation(question='What is 4 + 4?', context={}, metadata={}), Observation(question='What is 5 + 5?', context={}, metadata={})]
[Observation(question='What is 0 + 0?', context={}, metadata={}), Observation(question='What is 1 + 1?', context={}, metadata={})]
[Observation(question='What is 2 + 2?', context={}, metadata={}), Observation(question='What is 3 + 3?', context={}, metadata={})]

Since reset uses async_sample which returns the next datapoint in current dataset, we dont get deterministic results

If you think this issue is relevant I can create a new issue and contribute in it