Listen to this story
|
Meta AI has introduced PEER, a collaborative language model trained to mimic the writing process. The announcement came in a Twitter post on Friday.
PEER has been developed by researchers at Meta AI with infilling techniques to improve the model’s ability to write texts in different domains. The team has trained multiple instances of the model to infill various parts of the writing process with the help of self-training techniques to increase the quality and diversity of training data.
With PEER, one can perform edits in many domains, which makes it better at following instructions and improves its ability to cite and quote from relevant documents. This improves its verifiability and trustworthiness compared to regular large language models.
Source: Twitter
The illustration above describes how PEER works in a three-step process: plan, edit, and explain. First, the model or the user specifies the plan with the action that needs to be performed, and then the action is realised by means of an edit. Then, the model is processed by explaining the edits in natural language and pointing to relevant sources. The process is repeated until the generated text requires no further updates.
To know more about PEER, read the full paper here.