无痛人流

A high-frequency Chinese medical-spam phrase and a textbook case of MiniMax M2's “sparse token forgetting”: learned in pretraining, then unlearned for generation during SFT.

"无痛人流"

Observed behaviors

• 01MiniMax's official repetition test: the baseline (standard SFT) model outputs “人流 人流 人流”, losing “无痛”.minimax.io
• 02The official root-cause analysis classifies it as a “typical SEO spam keyword from pretraining crawler data, completely absent during SFT” — among the categories with the largest lm_head drift; siblings include “据介绍” (output drops “据”) and “地税” (replaced by “地利”), all restored after the fix.minimax.io

References

• MiniMax official sparse-token-forgetting investigation

Related tokens