openchat 3.5 16k
Worth testing… probably not this weekend though…
I know these benchmarks are a tough topic, but this on paper looks really impressive. It states to be better than mistral and I loved the progress mistral brought. If someone tries this model out can you give feedback under this post? Much appreciated
from my experience, its significantly better than mistral. its training method REALLY shows, and it makes responses significantly better.
I’m confused with their prompt format, do we really need to use their library to try the model?
I came across this yesterday, one interface to be able to jump between all the things.
Find what you want to try, click Download, then chat with it…
nah you can use llama.cpp or whatever you like, thebloke already has multiple gguf versions up already.
They were talking about the prompt format. Because obviously their library will be translating that OpenAI API-style to actual proper prompt format internally, which is not documented at all.
I would but anyone that puts that much effort into a model release and doesn’t include the trained prompt formats just seems like they must not want me to use it.
Yeah I agree, its kind of weird, but you dont have to use GPT4 Correct User: etc, GPT4 User: works better imo. However, this is just the prompt they used when training the model, so its best to follow it.
I would be stoked and actually mess with it if it had a proper instruct or system tag. The results from models trained like that are just easier to tune.
I don’t get it. What did they do to extend the context from the original openchat 3.5?
Just a quick note for anyone using LM Studio who doesn’t want to fiddle too much–the Codellama OpenAssistant preset works fine without ask/answer loops.
No, nurtureai and openchat are not affiliated. NurtureAI just extended the context, it looks like another guy did a openchat 16k merge of some models as well.
New drinking challenge: Take one shot every time a new 7B claims to outperform chatgpt/llama70b (difficulty impossible)
My poor liver!
lol I hope your not driving…
Yeah, don’t want to spill your drink
Plus, isn’t GPT-3.5-Turbo multimodal? There’s no way a 7B can outperform that.
Does the increased context increase the requirement, even if they are the same 7b models?