nemo_gym.vision_input

Shared prompt and image input construction for HLE datasets.

Module Contents

Functions

Name	Description
`_to_data_uri`	Normalize an HLE image field into a data URI usable as an `image_url`.
`build_image_input`	Build a materialized `responses_create_params.input` for one row.

API

nemo_gym.vision_input._to_data_uri(
    image: str
) -> str

Normalize an HLE image field into a data URI usable as an image_url.

The cais/hle image column stores a base64 data URI string (e.g. data:image/png;base64,...) for image questions. If a raw base64 payload is encountered instead, wrap it with a PNG data-URI prefix.

nemo_gym.vision_input.build_image_input(
    prompt_config: nemo_gym.prompt.PromptConfig,
    question: str,
    image: str
) -> list[dict]

Build a materialized responses_create_params.input for one row.

Applies the shared prompt template, then (for image questions) rewrites the user turn into a multimodal content list with an input_image block.