nemo_gym.web.datasets

Lossless upstream-row adapters for Gym web benchmark JSONL files.

Module Contents

Functions

Name	Description
`_start_urls`	-
`adapt_webvoyager_record`	-
`gym_row`	Wrap a model-independent task in the request shape consumed by WebAgent.
`load_json_records`	-
`write_jsonl`	-

API

nemo_gym.web.datasets._start_urls(
    value: typing.Any
) -> list[str]

nemo_gym.web.datasets.adapt_webvoyager_record(
    record: collections.abc.Mapping[str, typing.Any]
) -> dict[str, typing.Any]

nemo_gym.web.datasets.gym_row(
    task: nemo_gym.web.models.WebTask
) -> dict[str, typing.Any]

Wrap a model-independent task in the request shape consumed by WebAgent.

The policy adapter owns instructions, tools, image folding, and output parsing. Keeping those fields out of prepared data lets the exact same immutable 552-task population run with Nano Omni, Qwen, or another policy.

nemo_gym.web.datasets.load_json_records(
    path: str | pathlib.Path
) -> list[dict[str, typing.Any]]

nemo_gym.web.datasets.write_jsonl(
    rows: collections.abc.Iterable[collections.abc.Mapping[str, typing.Any]],
    output: str | pathlib.Path
) -> int