I noticed that in the e2e training data preparing process, it seems that no special tokens like **"system", "assistant" or "think" tokens are included in the e2e data. So you are just using the same data format as block-ep(PTQ) in the e2e(QAT) process? No try for these structural information?
I noticed that in the e2e training data preparing process, it seems that no special tokens like **"system", "assistant" or "think" tokens are included in the e2e data. So you are just using the same data format as block-ep(PTQ) in the e2e(QAT) process? No try for these structural information?