WizardLM is a fine tuned 7B LLaMA model. It uses a large number of instructions with different difficulties to follow the dialogue to fine tune. The novelty of this model is that it uses LLM to automatically generate training data.
The WizardLM model uses a new method called Evol Instruct (a new method that uses LLM generation of human beings to independently batch generate open instructions of various difficulty levels and technical ranges to improve the LLM ability) to train through 70k computer generated instructions. This method generates instructions with different difficulty levels.
Evol Inspect uses the following five operations to expand prompts:
-
Add Constraint
-
deepen
-
Concretization
-
Add reasoning steps
-
Complex input
These operations are sequentially applied to the initial instruction to make it more complex, and the reply is generated by LLM.