Accuracy comparison with wizard-vicuna-13b and stable-vicuna-13b?

#1
by anon7463435254 - opened

Thanks for all your contributions.

In your opinion, how does this model stack up in terms of accuracy compared to ggml versions of wizard-vicuna-13b and stable-vicuna-13b?

I made some tests and my observations :

pt4-x-vicuna-13B semms the worst of those 3.
wizard-vicuna-13b is the best but sometimes getting not straight answers then have to ask again.
stable-vicuna-13b you mean standard vicuna (vanila?)? - seems a bit better than pt4-x-vicuna-13B but worse from wizard-vicuna-13b

It does seem to be a bit worse at factual accuracy but its ability to play roles and do creative writing type tasks seems to be far better.

I made some tests and my observations :

pt4-x-vicuna-13B semms the worst of those 3.
wizard-vicuna-13b is the best but sometimes getting not straight answers then have to ask again.
stable-vicuna-13b you mean standard vicuna (vanila?)? - seems a bit better than pt4-x-vicuna-13B but worse from wizard-vicuna-13b

Are you using the proper format? It works best with Alpaca format, not Vicuna

Of course .
How I could use the wrong format ?
Vicuna q5_1 work properly with llama.cpp

Of course .
How I could use the wrong format ?
Vicuna q5_1 work properly with llama.cpp

There is a User/Assistant format for vicuna, and a

Instruction:

### Response:

format for alpaca. Ours uses the latter

Yep ... like apla a or alpaca-lora format
Nothing new :)

Sign up or log in to comment