Local-Novel-LLM-project/Ninja-v1-NSFW
Ninja-v1-NSFW is a 7 billion parameter Mistral-7B-v0.1 based large language model developed by Local-Novel-LLM-project, fine-tuned on a novel dataset. It is optimized for high-quality Japanese and English generation, featuring enhanced memory for long-context generation and the ability to generate NSFW content. This model is designed for creative writing and conversational applications requiring explicit content generation.
Loading preview...
Ninja-v1-NSFW: Mistral-7B Fine-tune for NSFW Content
Ninja-v1-NSFW is a 7 billion parameter Large Language Model (LLM) developed by Local-Novel-LLM-project, based on the Mistral-7B-v0.1 architecture. This model has been fine-tuned on a novel dataset, distinguishing it from the base Mistral model.
Key Capabilities & Features
- Mistral-7B Base: Built upon the Mistral-7B-v0.1 foundation.
- Multilingual Generation: Achieves high-quality generation in both Japanese and English.
- Enhanced Memory: Designed to maintain context and coherence even during long-context generation.
- NSFW Content Generation: Specifically capable of generating Not Safe For Work (NSFW) content.
- Instruction Format: Adopts the Vicuna prompt format, supporting multi-turn conversations.
Development & Merging
This model was created with support from the first LocalAI hackathon. Its development involved several methods, including the use of Chatvector for multiple models, simple linear merging of result models, domain and sentence enhancement with LORA, and context expansion. The merge recipe includes components from:
- WizardLM2: For high-quality multitasking.
- NousResearch/Yarn-Mistral-7b-128k: For 128k context window capabilities.
- Elizezen/Antler-7B: Specialized for novel writing.
- NTQAI/chatntq-ja-7b-v1.0: For high-quality Japanese specialization.
- Elizezen/LewdSniffyOtter-7B: For Japanese NSFW specialization.
Usage Considerations
Users should be aware that the training data may introduce biases, and generated sentences should be reviewed carefully. For context expansion with YaRN, trust_remote_code should be set to True. Long inferences may require significant memory, and inference with llamacpp is recommended over Transformers where possible.