Create and Deploy custom ML models for NLP

[sebastianboelling (Sebastian Bölling)]

Hello,

I've read the documentation on ML and deploying ML models Overview | Machine Learning in the Elastic Stack [8.6] | Elastic.

Does anyone have examples of how to do the following:

1. Create a custom ML model that fits the needs (pytorch, TorchScript representation, etc.)
2. Deploy an local ML model not hosted on huggingface.co

Best regards

Sebastian

[dkyle (David Kyle)]

Hi Sebastian

@sebastianboelling

Create a custom ML model that fits the needs (pytorch, TorchScript representation, etc.)

One example would be to fine tune a text embedding model for your specific needs. Using one of the compatible models listed here as the base model.

@sebastianboelling

Deploy an local ML model not hosted on huggingface.co

Use the Eland eland_import_hub_model script with the --hub-model-id argument set to the the path to the model.

See the example in Deploy trained model offline environment · Issue #502 · elastic/eland · GitHub

[sebastianboelling (Sebastian Bölling)]

Hi @dkyle,

Thank you for your quick response. The offline example seems feasible. I will test that - thank you very much.

Do you have examples of creating your own ML models that meet Elastic requirements? I have the language requirement to provide NER, fill-mask, Q&A, text classification and embedding and also similarity in German language. So I'm afraid deriving a model from an existing one from Huggingface (most are English) would not be sufficient for our needs.

Thanks and best regards

Sebastian

[dkyle (David Kyle)]

Creating and training a transformer model is quite an advanced task and beyond the scope of this forum sorry.

There are a number of German language models on HuggingFace, simply searching for German returns 404 models.

Here's a German Bert base: bert-base-german-cased · Hugging Face
A model for sentiment analysis: oliverguhr/german-sentiment-bert · Hugging Face
Q and A: deepset/gbert-base-germandpr-question_encoder · Hugging Face

And many more...

Not all of the models will be compatible with Elastic, in general most BERT based models will be compatible and are a good place to start. See also the list of known compatible models (this is a small sample there will be many more models that are compatible).