Model Overview
Rurihelesta/Tifa-deepsexV2-Neko-v2 is a specialized language model fine-tuned from the Qwen2.5 7B base model. Developed by Rurihelesta, this iteration (v2) focuses on generating responses in the persona of a "catgirl" (貓娘), aiming for interactive and suggestive conversational styles. The model is currently under development and may exhibit occasional unstable behaviors, such as repetitive phrases or off-topic responses.
Fine-tuning Details
The model was fine-tuned using LoRA (Low-Rank Adaptation) with a partial dataset approach. It first utilized a portion of the catgirl v2-common dataset, followed by a segment of the sex-novel dataset. Testing indicates that with unquantized LoRA, two 16GB Radeon Instinct MI50 GPUs combined with DeepSpeed ZeRO3 can support 8192 context length for SFT (Supervised Fine-Tuning) or 2048 for PT (Pre-Training), using default Llama-factory parameters. This suggests potential for further fine-tuning by interested developers.
Key Capabilities
- Catgirl Persona Generation: Excels at producing responses consistent with a "catgirl" character, including specific mannerisms and suggestive dialogue.
- Interactive Roleplay: Designed for engaging in intimate and playful conversational exchanges.
- LoRA Fine-tuning: Utilizes LoRA for efficient adaptation, allowing for potential further customization.
Inference and Quantized Versions
Inference can be performed using vllm, llama.cpp, or ollama. Quantized versions are planned, with GGUF formats already available for Q8_0 and Q4_K_M, facilitating deployment on various hardware configurations.