Unanswered

AI Builder - Custom model - retraining

(2) Share

Report

Posted on by itssujeeth

Is it possible to retrain a custom entity extraction model, using a modified dataset?

I've prepared a model for entity extraction with custom entities. Identified some issues with the data and trying to retrain by editing the CDS entity(Table in data verse). But while editing the model for retraining, I can only see initially loaded data, not the modified records.

Categories:

AI Builder

I have the same question (0)

All responses (10)

Answers (1)

Norah on at

Like (0)

Report

Hi

Thank you for trying out entity extraction. I'm sorry that right now we won't load latest data when you're retraining or editing a model. To use your latest data, please create a new model.

And as regard to load lasted data, we will add an item to our backlog. You can also share your thoughts about what do expect the behavior should be so it could meet your needs.

Thanks
Norah

Was this reply helpful? Yes No
itssujeeth 59 on at

Like (0)

Report

@Norah - Thanks for answering the first question
A subsequent question - Even though I flush out the old data and reload a new dataset, I can still see the old data coming up if I edit the published model. Does this mean the old CDS entities are retained in external storage?
If yes, any idea how long that data is retained?

Was this reply helpful? Yes No
Verified answer

Norah on at

Like (2)

Report

Yes, the dataset is stored in CDS, and will be deleted when the corresponding model is deleted.

Was this reply helpful? Yes No
michaelhoferat 14 on at

Like (7)

Report

Hi, what is the latest status concernign the backlog you mentioned. The reported problem is now more than 2 years ago. Is there any update in the meantime?

Creating a new model if new trining data is required is not a practicable solution.
We are using our models in several flows and creating a new model would require to change the model everywhere.
An other big disadvantage is that creating a new model also requires to tag the data in all trainig data datasets again. In our cases the data used for training is sometimes quite big. When then in practial usage some cases require a few more training datasets it is just not possible to do all that again.

Why is it not possible to just add some new datasets in dataverse, create a new version of a model, add the new data and tag it and retrain the model. So basically the same like when you are using models for extracting data from PDFs.

Was this reply helpful? Yes No
Aleclerc2013 7 on at

Like (4)

Report

Any update on this?

Was this reply helpful? Yes No
darylvogan 274 on at

Like (4)

Report

any update????

Was this reply helpful? Yes No
sj5345 9 on at

Like (3)

Report

Still no update?

Was this reply helpful? Yes No
Javiavan 30 on at

Like (3)

Report

I have this same problem. I find that I have to recreate everything again. Has this really not been resolved in 3 years? Was the roadmap lost along the way?
It's frustrating!!

Was this reply helpful? Yes No
vnair 28 on at

Like (1)

Report

After all the talk about AI making our routine lives easier, this is a big gap indeed.

Hard to believe that the only way is to re-create all the custom entities and re-do the whole model.

And Microsoft's answer is just bite it and move along.

Was this reply helpful? Yes No
Jacob12348771 22 on at

Like (0)

Report

Has there been any movement on this since it was initially added the backlog way back in 2020? This seems like a huge limitation to using a custom model in AI Builder. How can we look to improve the accuracy of a model over time if there is no ability to add to the initial data set? Retraining and tagging the entire data set each time is not viable.

Was this reply helpful? Yes No