Training a custom NER model using a pre-existing model (en_core_web_sm)

[Shakeeey]

I have a custom NER model that is created using train and test .spacy files. I have watched many tutorials and they all use spacy.blank("en") but I want to use the "en_core_web_sm" model and append my own custom entities to it. I have watched videos on how to do it on spacy v2 but nothing on v3 yet.

I have tried changing the config file and updating the [components.ner] to source = "en_core_web_sm" but I still don't get the pre-existing entities like (GPE, Date, Person, etc.) in my output.

Here is the link to the same question with maybe more details in a stackoverflow question: https://stackoverflow.com/questions/76459442/spacy-v3-custom-ner-model-training

[kadarakos]

Hey Shakeeey,

To be able to give you better help may I ask you to post the config files that you are using? You can also take a look at a related thread, not exactly your issue, but it might give you some useful information: #10674 (comment)

[Shakeeey]

Hey,

here is the link of my repo for this project: https://github.com/Shakeeey/spacy_project

the double.cfg is for combining those two models and the config.cfg is for creating my custom model.

Note: I used this tutorial to follow along to merge them together - https://github.com/explosion/projects/tree/v3/tutorials/ner_double

[kadarakos]

It seems like you are sourcing the ner component from the en_core_web_sm: https://github.com/Shakeeey/spacy_project/blob/main/config.cfg#L24, however, its corresponding tok2vec does not seem to be sourced: https://spacy.io/usage/processing-pipelines#sourced-components.

In general we do not recommend starting from a pre-trained ner model because its easy for it to forget what it learned when it sees new data. I recommend just training a fresh ner component, but using the en_core_web_lg vectors or if you have a enough resourced using a transformer backbone.

[kadarakos]

Oh sorry, I made a mistake the ner component actually has it's own tok2vec! Nevertheless, we do not recommend fine-tuning it.