Tifa-Deepsex-14b-CoT is a dedicated large language model deeply optimized based on Deepseek-R1-14B architecture, which achieves two core breakthroughs through multi-stage training and 128k ultra-long context support technology:
- Role Play Enhancement: The model adopts 671B parameter RL reward algorithm, which solves the pain point that the original model dialog character does not fit. By fusing 10K manual annotation + 100K generative SFT data, it can accurately generate multi-round conversations that match the character settings
- Long-form generation optimization: Incremental pre-training of 0.4T novel corpus and introduction of anti-duplication DPO mechanism, which improves the coherence of texts over 200,000 words by 52% and achieves an accuracy rate of 98% in the correction of Chinese-English mix-ups.
The model provides three quantization versions F16/Q8/Q4, and can achieve an inference speed of 18 tokens per second on RTX4090 graphics cards, which is one of the few long text authoring models that can be applied at a commercial level.
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 is a 14B parametric macromodel optimized for roleplaying and long story generation
