Tifa-Deepsex-14b-CoT: a large model that specializes in roleplaying and ultra-long fiction generation

General Introduction

Tifa-Deepsex-14b-CoT is a large model based on Deepseek-R1-14B deep optimization, focusing on role-playing, fictional text generation, and Chain of Thought (CoT) reasoning ability. Through multi-stage training and optimization, the model solves the original model's problems of insufficient coherence in long text generation and weak role-playing ability, which is especially suitable for creative scenarios that require long-range contextual correlation. By fusing high-quality datasets and incremental pre-training, the model significantly enhances the contextual association ability, reduces non-answers, and eliminates Chinese-English mixing and increases domain-specific vocabulary for better performance in role-playing and novel generation. In addition, the model supports 128k ultra-long contexts for scenarios requiring deep dialog and complex authoring.

This is a version of Deepseek-R1-14B that is deeply optimized for long-form fiction and roleplaying scenarios, and has a simple Android client available for download.

Function List

• Supports in-depth dialog for role-playing scenarios, generating responses that match the character's personality and background.
• Provide fictional text-generation skills to create coherent, long-form stories or plots.
• Chain of Thought (CoT) reasoning skills for scenarios requiring logical deduction and complex problem solving.
• Supports 128k ultra-long context to ensure high coherence and consistency of long text generation.
• The optimized model reduces the phenomenon of answer rejection, and security is moderately preserved for diverse authoring needs.
• Provide a variety of quantization versions (e.g., F16, Q8, Q4), adapting to different hardware environments and facilitating user deployment and use.

Using Help

Installation and Deployment

The Tifa-Deepsex-14b-CoT model is hosted on the Hugging Face platform, and users need to select the appropriate model version (e.g., F16, Q8, Q4) based on their hardware environment and requirements. Below is the detailed installation and deployment process:

1. Download model

• Visit the Hugging Face model page at https://huggingface.co/ValueFX9507/Tifa-Deepsex-14b-CoT.
• Select the appropriate quantization version (e.g. Q4_K_M.gguf) according to the hardware support. Click on the corresponding file to download the model weights.
• If you need to use the Demo APK, you can directly download the officially provided demo program (you need to manually import the character card and select the custom API).

2. Environmental preparation

• Ensure that the Python environment is installed (Python 3.8 or above is recommended).
• Install the necessary dependent libraries such as transformers, huggingface_hub, etc. They can be installed with the following commands:

  pip install transformers huggingface-hub
  

• If you are using a GGUF format model, it is recommended to install the llama.cpp or related support libraries. They can be cloned and compiled with the following commands:

  git clone https://github.com/ggerganov/llama.cpp
  cd llama.cpp
  make
  

3. Model loading

• Use transformers to load the model:

  from transformers import AutoModelForCausalLM, AutoTokenizer
  model_name = "ValueFX9507/Tifa-Deepsex-14b-CoT"
  tokenizer = AutoTokenizer.from_pretrained(model_name)
  model = AutoModelForCausalLM.from_pretrained(model_name)
  

• If the GGUF format is used, it can be run via llama.cpp:

  . /main -m Tifa-Deepsex-14b-CoT-Q4_K_M.gguf --color -c 4096 --temp 0.7 --repeat_penalty 1.1 -n -1 -p "your cue word"
  

  where -c 4096 can be adjusted to a larger context length (e.g. 128k) as needed, but be aware of hardware limitations.

4. Configuration and optimization

• Ensure that the returned context is stripped of think labels (e.g. ) to avoid affecting the model output. This can be achieved with the following code:

  content = msg.content.replace(/[\s\S]*? /gi, '')
  

• If you use the front-end interface, you need to manually modify the front-end code to adapt the context processing, refer to the official sample template.

Functional operation flow

role-playing feature

1. Enter the character setup: specify the character's background, personality, dialog scenes, etc. in the prompt. Example:

   You are a brave adventurer named Tifa who is exploring a mysterious ancient city. Describe your adventure and talk to the NPCs you encounter.
   

2. Generate Responses: The model generates personality-appropriate dialog or narratives based on characterization. The user can continue to input and the model will maintain contextual coherence.
3. Adjusting parameters: Optimize the output by adjusting temperature (to control the randomness of the generated text) and repeat_penalty (to control repeated content).

Novel Generation Function

1. Setting the story's context: provide the beginning or outline of the story, for example:

   In a distant kingdom, a young mage tries to unlock the secrets of time. Please continue this story.
   

2. Generate Story: The model will generate coherent long stories based on prompts, supporting multi-paragraph output.
3. Long context support: thanks to 128k context support, users can input longer story context and the model still maintains plot consistency.

Chain of Thought (CoT) reasoning

1. Enter complex problems: e.g:

   If a city generates 100 tons of waste per day, of which 60% is recyclable and 40% is non-recyclable, but the recycling facility can only handle 30 tons of recyclable waste per day, how will the remaining recyclable waste be handled?
   

2. Generate a reasoning process: the model analyzes the problem step by step, provides logical and clear answers, and supports long-range reasoning.

caveat

• Hardware Requirements: The model requires a high level of graphics memory to run, a GPU or high performance CPU with at least 16GB of graphics memory is recommended.
• Security and Compliance: The model retains certain security settings during training, and users need to ensure that the usage scenario complies with relevant laws and regulations.
• Context management: When using very long contexts, it is recommended to enter prompt words in segments to avoid exceeding hardware limits.

With these steps, users can easily get started with the Tifa-Deepsex-14b-CoT model, whether for role-playing, novel creation, or complex reasoning, and get high-quality generated results.