NuExtract language model

[juhoinkinen]

There exists models finetuned for structured extraction:

NuExtract-tiny, NuExtract, and NuExtract-large

We could try how they perform in comparison to currently used LMs.

[osma]

Looks very interesting, and it seems they have recently released new 1.5 models claimed to be even better: https://numind.ai/blog/nuextract-1-5---multilingual-infinite-context-still-small-and-better-than-gpt-4o

In practice, these are the new models:

• numind/NuExtract-1.5-tiny (based on Qwen2.5-0.5B, 0.5B params)
• numind/NuExtract-1.5 (based on Phi-3.5-mini-instruct, 3.8B params)

It appears they didn't release a 7B model in this round.

[osma]

I looked a bit at the usage code and prompt template that NuExtract uses. It seems to be based on "old-fashioned" completions, not chat-style interaction with messages. That's a bit unfortunate, because our fine-tuned models are based on the chat style, and going back doesn't seem like a very good idea...

Anyway, I guess we need to test it further.