kjr,
@kjr@babka.social avatar

I am trying to build a RAG with LLAMA 3 and... getting really crazy with the strange formats I get in the response....
Not only the response, but additional text, XML tags...

raf,
@raf@babka.social avatar

@kjr

Do you have a desired output format?

kjr,
@kjr@babka.social avatar

I realize now that maybe that is a question for @raf

kachumbali,
@kachumbali@mastodon.social avatar

@kjr tried Mistral7B instruct yet?

kjr,
@kjr@babka.social avatar

@kachumbali
Yes, it works fine, but only for English.

kachumbali,
@kachumbali@mastodon.social avatar

@kjr I’ll try to give it my first local RAG spin

kjr,
@kjr@babka.social avatar

@kachumbali If you want to try for English the results are quite good.
What is your use case?

kachumbali,
@kachumbali@mastodon.social avatar

@kjr I’m really not good at this. got Ollama running, tested it with local llms with a specific dataset, prompts, etc, cross evaluated responses, and would love to now improve that with a RAG setup and get better responses

kjr,
@kjr@babka.social avatar

@kachumbali I can recommend Mistral 7B, it has a very good performance.

  • All
  • Subscribed
  • Moderated
  • Favorites
  • llm
  • kavyap
  • thenastyranch
  • mdbf
  • DreamBathrooms
  • everett
  • magazineikmin
  • GTA5RPClips
  • Youngstown
  • cisconetworking
  • ethstaker
  • slotface
  • ngwrru68w68
  • rosin
  • cubers
  • JUstTest
  • InstantRegret
  • Durango
  • osvaldo12
  • modclub
  • tester
  • Leos
  • khanakhh
  • normalnudes
  • tacticalgear
  • megavids
  • anitta
  • provamag3
  • lostlight
  • All magazines