Dongxi NLP launches post-training series
Dongxi NLP has announced the launch of a post-training series. The first article points out that pretraining only teaches a model to "look at a piece of text and guess the next word," and that the training that turns a base model into a useful assistant after pretraining is collectively called post-training. The series divides post-training into three categories of techniques: instruction tuning, preference alignment, and RLVR, and will unfold over the coming weeks in the form of articles plus experimental code.
View original posthttps://t.co/z7WTfTRn20
View original
https://t.co/z7WTfTRn20
Body · AI translation
Pre-training teaches the model only one thing: given a piece of text, guess the next word.
After pre-training, all the training done to turn a base model into a useful assistant is collectively called post-training.
Post-training is usually divided into three techniques: instruction tuning, preference alignment, and RLVR.
This is the first article in the post-training series. In the coming weeks, we will strive to present post-training in an accessible way, accompanied by articles and experiment code.