-
Notifications
You must be signed in to change notification settings - Fork 12
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
SAGE Feedback & Future Development venues #13
Comments
Hi! I really like your models. I'm using the
|
Hi, @unterumarmung! FredT5-based models naturally inherit input / output lengths restrictions used while pre-training (it's 512 / 512, the details can be seen in the paper https://arxiv.org/pdf/2309.10931). |
Here is input text:
Here's the output text I get:
|
@unterumarmung thnx, we will be right back |
@unterumarmung we cut input length to 256 tokens while pre-training, because extensive text corruption added almost twice as many tokens to the source. The latter implies models are able to process sequence of 256 tokens in one run. Note however that you are still able to cut longer pieces of text into appropriate chunks of 256 tokens or less (split in sentences, for example). We didn't notice severe performance degradation when employing this approach to proceed with long texts. We expect this to work for you as well. Hope that helps and let us know if you experience noticeable drop in performance, when running longer text in chunks. |
Hi, everyone!
SAGE🌿 Team speaking.
🚀 This issue is a dedicated place for the organised feedback on our project and ideas on how we can improve SAGE to build best open source spellchecker for multiple languages.
We would love to hear from you about your positive and negative experiences with SAGE, your expectations and ideas. But please be concise and back up your feedback with representative and reproducible examples.
See you in comments below ⬇️
The text was updated successfully, but these errors were encountered: