Are there any plans to adopt the JEV model for KG tasks such as entity alignment?

[plpycoin]

LLMs are simply too costly and slow for extraction tasks...

[WaylandYang]

Thank you for raising this. It is an interesting direction and we would like to try it: extraction and entity alignment are where most of the model cost goes today, and a smaller model that does those two steps well would change what a large base costs to build.

Pull requests are welcome. Since this replaces a model call in the pipeline, please open an issue first saying which step you would start with (extraction, kind-word alignment, phrase alignment or entity adjudication) and how you would measure it; the benches under scripts/bench give a before and after for each of those steps.

Thanks again for the idea.

[plpycoin]

• This illustration from Daily Dose captures the approach I want to implement. Once I finish what I'm working on, I'll set aside a solid block of time to bring in JEV.