Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Make the T5 model tests use cosine similarity #895

Merged
merged 4 commits into from
Feb 10, 2025

Conversation

sogartar
Copy link
Contributor

@sogartar sogartar commented Feb 1, 2025

There were several xfail tests with bad metric. Cosine similarity is a better metric for language embeddings.

The comparison between bf16 and f32 exhibits a small fraction of outliers that have a higher per-token numerical error than the majority of tokens. To account for that the testing metric is expanded to test for inlier and outlier absolute tolerance.

@sogartar sogartar marked this pull request as ready for review February 3, 2025 14:44
@ScottTodd
Copy link
Member

There were several xfail tests with bad metric. Cosine similarity is a
better metric for language embeddings.

The comparison between bf16 and f32 exhibits a small fraction of
outliers that have a higher per-token numerical error than the majority
of tokens. To account for that the testing metric is expanded to test
for inlier and outlier absolute tolerance.
@sogartar sogartar merged commit 7a8f360 into nod-ai:main Feb 10, 2025
34 checks passed
monorimet pushed a commit that referenced this pull request Feb 13, 2025
There were several xfail tests with bad metric. Cosine similarity is a
better metric for language embeddings.

The comparison between bf16 and f32 exhibits a small fraction of
outliers that have a higher per-token numerical error than the majority
of tokens. To account for that the testing metric is expanded to test
for inlier and outlier absolute tolerance.
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