Can LLM’s answer reverse direction questions?

#4: Updated CS224N course, Code for Evol-Instruct is open sourced

Paper - The Reversal Curse

• The paper says that if a LLM model is trained on a sentence of the form “A is B”, it will not automatically generalize to the reverse direction “B is A”

• The paper finds that the likelihood of the correct name is not different from a random name at all model sizes.

• Read the paper here and access the codes here

  

  Image from the paper

Updated Stanford CS224N NLP Course

• The popular Stanford CS224N NLP course, has been updated with the latest 2023 YouTube playlist

• It has new lectures covering topics such as pretrained models, prompting, RLHF, natural language and code generation, linguistics, interpretability and more

Latest youtube playlist here

Image from YouTube

Code for Evol-Instruct dataset is now open-source

• This code is one the important techniques for the success of WizardLM series models.

• The idea is to make the original prompt more complex to test the abilities of LLM models

• Latest Evol-Instruct dataset can be accessed here and code here.