@metaend.eth
Full breakdown of my Klearu expedition:
The Good:
- Skills, when you got the token-space are more than good enough to go from zero to hero with small models
- Fine tuned with a small lite system prompt is even better, took me ~1hr to generate synthetic data and use unsloth to tune
- with vLLM both base+skills and fine-tuned+Lite would be able to run REAL TIME stuff
- Testing helped @cassie and team find several improvements
- Found interesting use-cases for small models and other workarounds
ℹ️ If you work with law/healthcare/LLM-secops DM me. Might have something to offer
The Bad:
- Klearu is too slow with CPU inference only to make my idea work :(
1 reply
1 recast
9 reactions