Find answers from the community

Updated 6 months ago

@Logan M I am trying to create a graph

@Logan M I am trying to create a graph with Llama 3 70b from Bedrock but I see that a nessasary function acomplete isn't implemented? what is the best way to go getting the implemented, tested and PRed?
L
C
15 comments
probably need to edit the class and use an async boto3 session to send the requests. It will probably be mostly the same as the other methods, but with async and await

In terms of PRs, feel free to fork the repo, make changes, and open a PR πŸ™
are there test cases?
yes, since each inegration is its own package, there are unit tests for each

However, with stuff like LLM calls, a lot of it needs to be mocked, which takes some amount of work. I don't think bedrock has much testing at the moment due to that
Also do you know if anyone has done finetunes for llama 3 for entity extraction?
Hmm I do not know actually. But for entity extraction, using an LLM is overkill. Check out gliner for example
much smaller, much more scalable
Is there a way to use this with the property graph?
I was planning on finetuning llama 8B to get something that was fast/cheap enough to use in production ahaha h
this might be way better
Maybe! I've been thinking about it.

I think there's a few ways you could implement as a custom extractor
  • extracting entities, and then asking the llm to generate relationships
  • gliner has a new multitask model that let's you give an entity + relationship, and it predicts the matching entity to complete the triple (if any). If you had a list of predefined relationships types, you could run every pair
Personally I'm hoping for a true zero-shot model similar to gliner for triples. It's soooo close
I guess to rephrase my question we are creating a property graph and want to increase the efficiency so we could create the property graph. I need to look at the multitask model because simply labelling wouldn't create the nodes right? it would only tell you what nodes should exist?
Maybe could be useful in combination with a small LLM where you ask the small LLM to actually create the entity + relationship which is what you're saying in your first bullet point.
Looking into the multi-task model definitely seems more straight forward
Add a reply
Sign up and join the conversation on Discord