I've been mixed on this, it seems great for throwaway projects but a dead end for ones that survive:
* Out-of-the-box fine-tuned model for extraction makes for an easier start, so happy to see this out there
* ... But presumably one of the next steps in a project using it is doing further fine-tuning on their data after some use, but this is a closed model with closed data, so preserving the original fine tuning quality is unclear without that
Hi there, CEO of NuMind here, absolutely agree, these models are meant to be fine-tuned to be really useful. Right now you can see how the input -> output are formatted in the model card on HuggingFace. From that you can create a set of input -> output and fine-tune NuExtract. We are working on a platform to make this fine-tuning easier.
Where is the fine-tuning source code + data and with what license? Afaict that's the proprietary part which is why it isn't in their github and hf doesn't link to it.
Hence my comment that I don't see how to do further fine-tuning that doesn't just wipe out what this does. They did a free binary release of fine-tuned weights afaict, which is problematic for building a production data pipeline that you re-fine-tune over time.
I see. Yeah, I'd be interested in fine tuning as well on my own extractions. The tiny model even says it's meant to be fine tuned. I guess that's what they are commercializing :/
It's for NER (Named Entity Extraction). In the past, models had to be trained to know how to find entities such as names, addresses, dates. This model has the capability to do that zero-shot.
In other words, you can pass it text then say extract the names, addresses and dates.
* Out-of-the-box fine-tuned model for extraction makes for an easier start, so happy to see this out there
* ... But presumably one of the next steps in a project using it is doing further fine-tuning on their data after some use, but this is a closed model with closed data, so preserving the original fine tuning quality is unclear without that
Hence my comment that I don't see how to do further fine-tuning that doesn't just wipe out what this does. They did a free binary release of fine-tuned weights afaict, which is problematic for building a production data pipeline that you re-fine-tune over time.
In other words, you can pass it text then say extract the names, addresses and dates.