TARS is trained on high-quality tokens of billions in vertical industries, fully reproducing and improving the pre train, SFT, and RLHF stages of large-scale language model training, and achieving good results in language understanding, instruction following, and other aspects. In addition, in pursuit of higher security standards, the TARS model further integrates multiple cutting-edge NLP technologies such as inappropriate speech discrimination, helping enterprises take a step forward in the digital and intelligent transformation wave.