WizardCoder 34B: Complex Fine-Tuning Explained

Significant performance difference of WizardCoder-Python-34B-V1.0 (based on Code LLama 2) explained, when compared to simple Code LLama - Python 34B.
The difference is the added complexity cascade in the evolving instruction fine-tuning.
original source (all rights with authors):
github.com/nlpxucan/WizardLM
#agents
#llama2
#codegeneration

[@Nick_With_A_Stick]

This video is just perfect. Literally exactly what I needed thanks for your hard work! when I make my datasets I use a evol and orca with human guided prompting technique. It tends to work pretty well.

[@cecilsalas8721]

Excellent content ! 😊

[@yannickpezeu3419]

Thanks !

[@ashwinkotgire1372]

Can you please tell me whether the Instruct finetuning was done on Llama-chat or normal Llama model? And the GPT model used to generate the data was ChatGPT or normal GPT provided by OpenAi?

[@fernandofernandesneto7238]

Fantastic explanation. By the way, the langchain llm does not exhibit critic / reasoning capabilities ... that's what would be missing (am I right?)