Skip to content

Conversation

@hinthornw
Copy link
Collaborator

@hinthornw hinthornw commented Jul 1, 2023

  • change to ABC - this lets us add things like the evaluation name for loading

@vercel
Copy link

vercel bot commented Jul 1, 2023

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment
Name Status Preview Comments Updated (UTC)
langchain ⬜️ Ignored (Inspect) Jul 6, 2023 7:48am

@vowelparrot vowelparrot force-pushed the wfh/chain branch 4 times, most recently from 99e867b to f3de5c4 Compare July 3, 2023 18:49
@hinthornw hinthornw requested review from agola11 and hwchase17 July 3, 2023 19:51
@agola11
Copy link
Collaborator

agola11 commented Jul 6, 2023

looks like this needs to be stacked ontop of #6942

@hinthornw
Copy link
Collaborator Author

looks like this needs to be stacked ontop of #6942

My fault - should be other way around actually (#6942 stacks on this one)

If a string evaluator doesn't want references, it currently silently
ignores any that are passed in. This PR proposes to warn once if they're
provided. It does so via a mixin.

Other options we could do:
- Separate classes for criteria with references and comparisons with
references - I don't love this because all the combinations lead to
class explosion and the class names themselves are ugly
(CriteriaWithReferencesNoLabels, etc.)
- Raise errors if they're provided - I don't like this because it
requires more checks when running evaluators over a dataset, but would
be more explicit.
- Decide at calling time -> could be a lot of partial prompt insertions
that makes it more brittle. Would make a single "run" potentially have
two results that are fundamentally different evaluations
@hinthornw hinthornw merged commit ec66d51 into master Jul 6, 2023
@hinthornw hinthornw deleted the wfh/chain branch July 6, 2023 13:37
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants