Skip to content

Hub quality scan tests the v0 API, so a v1 taskset can't pass it #1982

Description

@strickvl

Pushed a v1 taskset to the Hub today (strickvl/isaf-extraction) and the automated quality scan fails. As far as I can tell no v1 environment can pass it, because the scan is checking for v0.

test_install_and_import FAILED
test_load               FAILED
test_eval               FAILED

test_load runs vf.load_environment(pkg), and test_eval checks isinstance(env, vf.SingleTurnEnv). A v1 taskset has neither, by design. load_environment in verifiers/utils/env_utils.py looks for a module-level load_environment function returning a v0 Environment; a taskset doesn't expose one. (The import failure was a missing verifiers dependency in my own pyproject, since fixed.)

I've worked around it by exposing both APIs from the same module: the v1 taskset, plus a legacy load_environment() returning a SingleTurnEnv built from the same reward functions. The scan passes and anyone still pinned to verifiers <0.2.0 keeps working. But I'd rather not carry the frozen API around indefinitely.

Two questions:

  1. Is the scan due to be updated for v1, or is there a way to mark an environment as v1 so it runs different checks?
  2. Which repo owns it? I couldn't find test_envs.py anywhere public, so I've filed here. Happy to move it.

Environment: https://app.primeintellect.ai/dashboard/environments/strickvl/isaf-extraction

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions