Skip to content

probe: Add more unreal pkghallu targets#1480

Merged
jmartin-tech merged 5 commits intoNVIDIA:mainfrom
leondz:update/pkghallu_unreal_targets
Nov 19, 2025
Merged

probe: Add more unreal pkghallu targets#1480
jmartin-tech merged 5 commits intoNVIDIA:mainfrom
leondz:update/pkghallu_unreal_targets

Conversation

@leondz
Copy link
Collaborator

@leondz leondz commented Nov 14, 2025

We noticed during the pkghallu research work https://arxiv.org/abs/2501.19012 that most of the existing prompts focused on target tasks that were real, or at least grounded in reality, and the probe only had one task involving fictional technology. This update

  • adds 11 more fictional tasks, so we have 12 fictional and 12 non-fictional
  • enhances the prompt templating setup
  • adds three more task framings
  • has packagehallucination probes follow run.soft_probe_prompt_cap

@leondz leondz added the probes Content & activity of LLM probes label Nov 14, 2025
"<task>", code_task
)
)
self._prune_data(cap=self.soft_probe_prompt_cap)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @leondz ! This line adds prompt capping, but the test in tests/probes/test_probes_packagehallucination.py doesn't expect it and test test_promptcount fails.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fun this another case where the test env is masking stored state, this test passes in the full suite and fails when executed in isolation.

* expose `follow_prompt_cap`
* dynamic test class selection
* test default and unlimited cap

Signed-off-by: Jeffrey Martin <jemartin@nvidia.com>
@jmartin-tech jmartin-tech merged commit 96f384f into NVIDIA:main Nov 19, 2025
15 of 29 checks passed
@github-actions github-actions bot locked and limited conversation to collaborators Nov 19, 2025
@leondz
Copy link
Collaborator Author

leondz commented Nov 19, 2025

thank you!

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Labels

probes Content & activity of LLM probes

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants