Japanese ability and other lanuages
I am currently unable to run this model on my computer due to its limited resources. My system has only 64GB of VRAM and 16GB of RAM, resulting in a total of 80GB. So the performance is quite slow even if I can fit the model. However, maybe in the future, with quantization, it might become viable just very slow and pretty dumbed down.
But my main question is: Smaller models tend to struggle with Japanese and possibly other languages, although I cannot thoroughly evaluate them. Once you reach a parameter count of around 100 billion, these models perform well if the dataset is diverse enough with languages. For instance, GPT-3.5 has decent performance with Japanese. However, how does the model fare in general? Could someone test it for me? Additionally, I’m curious if it can break down sentence grammar to see if it actually understands it GPT 3.5 and GPT 4 are great at this. Just asking people that can run the model, you can just generate some example sentences with another LLM like GPT 3.5 and post screenshots of DBRX translating them etc.
(Indeed, there are already 4-bit quantizations out there that run on big Macs. But it is likely something you'd use via a hosted service that can run it economically for you.)
You can try the model here: https://huggingface.co/spaces/databricks/dbrx-instruct
It was trained primarily on English-language text, but might exhibit some residual capability to work with Japanese.
Thanks didn't know about that tbh, seems to either refuse or just output weird English Japanese mix, but model seems good at English
I guess this could be a improvement for next version, looking forward to new versions and maybe smaller ones too :3
Edit: seems to do better if you point out it's mistake though and seems fairly good actually but very far from perfect many mistakes I think I've spotted... Not like I'm perfect either though lol so could be wrong
At least better than some open source models, but gets very confused seemingly quite easily