nemo_gym.web.models

Wire models for stateful, multimodal web environments.

Module Contents

Classes

Name	Description
`WebAction`	Model-independent computer-use action accepted by visual browsers.
`WebActionLimits`	Runtime-owned waiting limits advertised when a browser session starts.
`WebActionProfile`	-
`WebArtifactRef`	Stable reference to an artifact retained outside a model request.
`WebBenchmark`	-
`WebImage`	An image can be transported inline, retained as an artifact, or both.
`WebObservation`	JSON-safe observation sent from the browser runtime to the agent.
`WebObservationProfile`	-
`WebRuntimeProfile`	Browser execution protocol exposed by Gym.
`WebStepResult`	-
`WebTab`	-
`WebTask`	Normalized task envelope; original benchmark fields remain lossless.
`WebVerifierResult`	-

Data

BROWSER_TARGET_CLOSED_STATUS

CAPTCHA_BUDGET_EXHAUSTED_STATUS

API

class nemo_gym.web.models.WebAction()

Bases: BaseModel

Model-independent computer-use action accepted by visual browsers.

answer

Optional[str] = None

arguments

dict[str, Any] = Field(default_factory=dict)

metadata

dict[str, Any] = Field(default_factory=dict)

name

str

raw_model_output

str = ''

script

str

terminal

bool = False

class nemo_gym.web.models.WebActionLimits()

Bases: BaseModel

Runtime-owned waiting limits advertised when a browser session starts.

The default retains the historical tool contract. A benchmark profile may permit longer waits without rewriting the model’s requested duration.

max_step_wait_seconds

float | None

max_wait_seconds

float

nemo_gym.web.models.WebActionLimits.validate_wait_budget() -> 'WebActionLimits'

class nemo_gym.web.models.WebActionProfile

Bases: enum.Enum

COMPUTER_USE

= 'computer_use'

class nemo_gym.web.models.WebArtifactRef()

Bases: BaseModel

Stable reference to an artifact retained outside a model request.

mime_type

str

sha256

str = Field(pattern='^[0-9a-f]{64}$')

size_bytes

int = Field(ge=0)

uri

str

class nemo_gym.web.models.WebBenchmark

Bases: enum.Enum

VISUALWEBARENA

= 'visualwebarena'

WEBARENA

= 'webarena'

WEBVOYAGER

= 'webvoyager'

class nemo_gym.web.models.WebImage()

Bases: BaseModel

An image can be transported inline, retained as an artifact, or both.

artifact

Optional[WebArtifactRef] = None

data_url

Optional[str] = None

nemo_gym.web.models.WebImage.require_transport() -> 'WebImage'

class nemo_gym.web.models.WebObservation()

Bases: BaseModel

JSON-safe observation sent from the browser runtime to the agent.

active_tab_index

int = Field(default=0, ge=0)

axtree_text

str = ''

elapsed_time

float = Field(default=0.0, ge=0.0)

element_map

dict[str, dict[str, Any]] = Field(default_factory=dict)

focused_element_id

str = ''

goal

list[dict[str, Any]] = Field(default_factory=list)

last_action

str = ''

last_action_error

str = ''

metadata

dict[str, Any] = Field(default_factory=dict)

screenshot

Optional[WebImage] = None

tabs

list[WebTab] = Field(default_factory=list)

url

str = ''

class nemo_gym.web.models.WebObservationProfile

Bases: enum.Enum

A11Y

= 'a11y'

SCREENSHOT

= 'screenshot'

SOM

= 'som'

class nemo_gym.web.models.WebRuntimeProfile

Bases: enum.Enum

Browser execution protocol exposed by Gym.

Model-specific prompting is deliberately not represented here. Nano Omni, Qwen, and future policies all drive the same visual browser runtime after their policy adapters normalize actions into WebAction.

VISUAL_BROWSER

= 'visual_browser'

class nemo_gym.web.models.WebStepResult()

Bases: BaseModel

benchmark_reward

Optional[float] = None

execution_ok

bool

info

dict[str, Any] = Field(default_factory=dict)

observation

WebObservation

terminated

bool = False

truncated

bool = False

class nemo_gym.web.models.WebTab()

Bases: BaseModel

active

bool = False

index

int = Field(ge=0)

title

str = ''

url

str = ''

class nemo_gym.web.models.WebTask()

Bases: BaseModel

Normalized task envelope; original benchmark fields remain lossless.

action_profile

WebActionProfile = WebActionProfile.COMPUTER_USE

auth_profile

Optional[str] = None

benchmark

WebBenchmark

input_images

list[str] = Field(default_factory=list)

intent

str = ''

model_config

= ConfigDict(extra='allow')

observation_profile

Optional[WebObservationProfile] = WebObservationProfile.SCREENSHOT

original_metadata

dict[str, Any] = Field(default_factory=dict)

runtime_profile

WebRuntimeProfile = WebRuntimeProfile.VISUAL_BROWSER

seed

int = 0

sites

list[str] = Field(default_factory=list)

start_urls

list[str] = Field(default_factory=list)

task_id

str

task_kwargs

dict[str, Any] = Field(default_factory=dict)

verifier_profile

Optional[str] = None

nemo_gym.web.models.WebTask.stringify_task_id(
    value: typing.Any
) -> str

classmethod

class nemo_gym.web.models.WebVerifierResult()

Bases: BaseModel

evidence

list[WebArtifactRef] = Field(default_factory=list)

failure_kind

Optional[str] = None

metadata

dict[str, Any] = Field(default_factory=dict)

raw_score

float = 0.0

reward

float = 0.0

subscores

dict[str, float] = Field(default_factory=dict)

task_success

bool = False

valid_sample

bool = True

verifier_version

str = ''

nemo_gym.web.models.BROWSER_TARGET_CLOSED_STATUS = 'browser_target_closed'

nemo_gym.web.models.CAPTCHA_BUDGET_EXHAUSTED_STATUS = 'captcha_budget_exhausted'