base_model:
- ValueFX9507/Tifa-Deepsex-14b-CoT
language:
- zh
- en
library_name: transformers
tags:
- incremental-pretraining
- sft
- reinforcement-learning
- roleplay
- cot
- sex
- SFW
license: apache-2.0
Tifa-Deepseek-14b-CoT-Chat
A huggingface format conversion of the GGUF from here: https://huggingface.co/ValueFX9507/Tifa-Deepsex-14b-CoT
For merging, requantizing, finetuning and such. A partial translation of the model card:
Standard data training, mature RL strategy, additional anti-duplicate reinforcement learning, suitable for normal use, normal output text quality, and divergent thinking in a few cases.
Incremental training of 0.4T novel content
100K SFT data generated by TifaMax, 10K SFT data generated by DeepseekR1, 2K high-quality artificial data
30K DPO reinforcement learning data generated by TifaMax to prevent duplication, enhance contextual association, and improve political security
16k ultra-long context training
Random truncation training enhances robustness
8×H20 GPU full-scale fine-tuning
Personal observations:
Don't let the DeepSex name fool you. This model is strong at SFW, English, long form (>32K context) storywriting, especially for a 14B, with good comprehension of the whole plot, details and the current state of the story. This is interesting, as it was "only" trained at 16K and (seemingly) mostly in Chinese.
Subjectively, the "crazy" version feels a little stronger, hence I am mostly testing with that.