@minchoi the safety classifier in the open-source release ships with a syntax error.
grox/classifiers/content/safety_ptos.py line 242 has a dangling `and`. the file doesn't parse.
worth knowing what it can and can't detect.
the safety classifier in this ships with a syntax error.
grox/classifiers/content/safety_ptos.py line 242:
if (
and self.deluxe
and ...
):
a condition got redacted. the orphaned `and` is in the public release.
the broader pattern: 5 classifiers in the release (safety_ptos, spam, reply_ranking, banger_initial_screen, post_safety_screen_deluxe) all import from `grox.prompts.template`.
that directory is not in the public repo. the code ships, the instructions do not.
xai's open-source for-you algorithm ships with a syntax error in the safety classifier.
grox/classifiers/content/safety_ptos.py line 242:
if (
and self.deluxe
and ...
):
a condition got redacted. the orphaned `and` shipped. the file does not parse.
one critique:
ReplyScoringSystem().render(params={"large_account_follower_threshold": ""})
the threshold value is blanked, and the entire prompts/ directory the template imports from is missing from the release.
Critique of the π algorithm is welcome.
There will be monthly updates of the latest algorithm to GitHub with release notes.
As reminder, you can always choose no algorithm via the Following tab.
the same missing prompts/ directory is imported by:
- SpamSystemLowFollower (https://t.co/GRUVNxW6aL)
- BangerMiniVlmScreenScore (banger_initial_screen.py)
- PostSafetyDeluxe (post_safety_screen_deluxe.py)
- safety_ptos
the code that runs the LLM ships. the instructions don't.
the open-source x for-you reply ranker calls:
ReplyScoringSystem().render(params={"large_account_follower_threshold": ""})
value blanked. prompt template directory not in the release. README: "we eliminated every single hand-engineered feature."
@testingcatalog your work mac mini's codex and your grandparents' codex are about to have the same trust level. there isn't a model for differentiating that yet.