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.

[WaylandYang]

Thank you for following up, and for sharing the illustration. Looking forward to it whenever you have the time; there's no rush.

When you start, please open an issue first: say which step you'd begin with (extraction, kind-word alignment, phrase alignment or entity adjudication) and how you'd measure it against the benches in scripts/bench.

One thing worth knowing: we're planning an overall rework of extraction in the near term, and where a smaller model like this could fit will be considered as part of it. Checking in on the issue before writing code will help the two line up.