A complete program for improving the coherence of long plots
Tifa-Deepsex-14b-CoT, although it supports 128k contexts, still requires specific techniques to ensure the quality of very long texts in practice:
- staged writing approach::
- First generate a 500-word outline (example prompt: "Please create a detailed outline for the [title] novel that includes 5 key turning points")
- Chapter-by-chapter generation includes a summary of the previous article in the cue word (e.g., "Pick up the following plot: [3-sentence summary], detailing the protagonist's encounters as he enters the tomb.")
- Hardware Adaptation Tips::
- When using the Q4_K_M quantization version, set the context length to 8192 instead of the maximum value to avoid performance degradation
- Add the transformers when they are loaded
max_lengthParameters:model.generate(input_ids, max_length=8192, temperature=0.8)
- Post-calibration tools: In conjunction with the use of llama.cpp's
--ctx-mmapparameter to implement long text memory mapping, or develop customized coherence checking scripts to detect character name/timeline/location coherence
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 " How to optimize Tifa-Deepsex-14b-CoT for plot coherence in long story generation?
