Code Generation SuperCoT
I wonder if u know that this is the best model right now ?? this little boy of 30B's codes better than GPT4..
CoT has always been the secret.. SuperCoT has a real difference. We need to get metrics of this and larger context.. potentially looking at a gamechanger dataset "SuperCoT". Gonna try to FT something smaller and see what comes out..
Looking forward for a 4k, 8k ctx models!!
Great work bro, u need to work on the README and bring more data on this, very very promising.
You might consider giving your feedback here: https://huggingface.co/kaiokendev/SuperCOT-LoRA
All I did was quantize a LoRA merge someone else made, I didn't do anything with training the LoRA used to make this. That link above is the SuperCOT LoRA which is responsible for how this model performs. Their readme has more information on what's included and how it was made.
Well... I guess I wasn't wrong :D