2 ms·
Hopefully this update also comes to ollama soon <3 So far I have to say I prefer llama3 for general interaction. It feels more natural. And this really is from
by wkat4242 2y ago
Hopefully this update also comes to ollama soon <3
So far I have to say I prefer llama3 for general interaction. It feels more natural. And this really is from all the big tech ones: llama3, phi3, gemma2, mistral, aya..
I really miss a bigger version of llama3 though. 8b is not using my gpu effectively. I could easily fit a 14b version but there isn't one. The next step up 70b is way too big.
- marcyb5st 2y agoWeird, I feel Gemma2 is overall better even if it doesn't follow instructions a bit more often than llama3. Luckily filtering out preambles like "Sure, let me help you with that" from primarily JSON/YAML answers is easy enough IMHO.
- Filligree 2y agoYou can use a grammar to enforce the output being valid JSON.
- luke-stanley 2y agoDo you know any tools that enforce grammar or structure, that allow the use of general instructions (e.g., use 5 positive English words)? I believe I saw such a tool mentioned by Abetlen, although it wasn't their tool, but I can't seem to find it now. I've already checked LMQL, SGLang, Guidance, and Outlines. Thanks.
- luke-stanley 2y agoAnswering my own question: Another project called Instructor has a `llm_validator` method that fits in with it's Pydantic oriented constrained typed generation. https://python.useinstructor.com/api/#instructor.dsl.validators.llm_validator https://python.useinstructor.com/api/#instructor.dsl.validat...