Meta launches PEER, a language model that is far superior to large editing models
Meta AI introduced PEER, a collaborative language model trained to mimic the writing process. The announcement came in a Twitter post on Friday.
PEER was developed by researchers at Meta AI with fill-in methods to improve the model’s ability to write texts in different domains. The team trained several examples of the model to fill in different parts of the writing process using self-training methods to improve the quality and variety of the training data.
PEER can be edited in many areas, making it better at following instructions and improving its ability to cite and quote relevant documents. This improves its verifiability and reliability over conventional large language models.
The illustration above describes how PEER works in a three-step process: planning, editing, and explaining. First, the model or user specifies a plan with an action to be performed, and then the action is implemented through editing.
The model is then processed by explaining the edit in natural language and specifying the appropriate sources. The process is repeated until the generated text needs further updating.