As the stabilized version of the Tifa series, Deepsex-14b-CoT-Chat features an industry-leading triple protection design:
- data filtering: The training set was subjected to two rounds of content auditing by Claude-4 and GPT-4, with a harmful content filtering rate of 99.7%
- RLHF Enhancement:: 30K DPO data is specifically aligned to politically sensitive topics, with breach response rates kept below 0.31 TP3T
- Dynamic monitoring: Integration of neural network guard module in the output layer to detect and block non-compliant generated content in real time
While retaining the ability to create a 0.4T high-quality fictional corpus, the role-playing vividness remains at the top of the industry at 89 out of 100 while achieving an S rating in the safety test.
This answer comes from the articleTifa-Deepsex-14b-CoT: a large model that specializes in roleplaying and ultra-long fiction generationThe
May not be reproduced without permission:AI productivity tools " Tifa-Deepsex-14b-CoT-Chat version balances creativity and security
