Promising stuff from their repo, claiming “exceptional performance, achieving a [HumanEval] pass@1 score of 57.3, surpassing the open-source SOTA by approximately 20 points.”
From the Twitter post
New StarCoder coding model from @WizardLM_AI
“WizardCoder-15B-v1.0 model achieves 57.3 pass@1 on the HumanEval Benchmarks … 22.3 points higher than the SOTA open-source Code LLMs.”
My quants: https://huggingface.co/TheBloke/WizardCoder-15B-1.0-GGML https://huggingface.co/TheBloke/WizardCoder-15B-1.0-GPTQ
Original: huggingface.co WizardLM/WizardCoder-15B-V1.0 · Hugging Face
11:21 AM · Jun 14, 2023
On The Bloke’s hugging face repo, it says the GGML quants are not compatible with llama.cpp, anyone know why?
It’s a different type of model. llama.cpp only supports LLaMA models while GGML (the machine learning library llama.cpp is based on) has examples of various models with different architectures. WizardCoder, MPT, Bloom, probably very soon Falcon. Also some separate projects use GGML to support other models (including some of the ones I listed). For example the Rust “llm” project can support LLaMA models, MPT, BLOOM.
So if I understand correctly it is fine tuned for coding or what exactly is this Wizard model doing?
It is StarCoder and fine tuned on a new Wizard Instruct dataset optimized for coding models. So it follows the instruct formatting of prompts on top of the StarCoder base model.
That sounds great honestly! Does that work with the newest ggml yet?
Doesn’t look like it, hopefully it does someday. I am stoked to try this one out.