CyberpunkLegend/Qwen2.5-7B-Instruct-CharacterEnhance
CyberpunkLegend/Qwen2.5-7B-Instruct-CharacterEnhance is a 7.6 billion parameter QLoRA fine-tuned model based on Qwen2.5-7B-Instruct, specifically optimized for bilingual (English and Chinese) character role-play dialogue generation. It excels at producing natural, character-consistent responses in conversational AI scenarios. The model was trained on the PIPPA dataset to enhance its ability to maintain character personas.
Loading preview...
Model Overview
CyberpunkLegend/Qwen2.5-7B-Instruct-CharacterEnhance is a 7.6 billion parameter language model, fine-tuned using QLoRA (4-bit NF4 quantization) on the robust Qwen2.5-7B-Instruct base. Its primary focus is bilingual character role-play dialogue generation, supporting both English and Chinese.
Key Capabilities
- Character-Consistent Dialogue: Generates natural and in-character responses for role-play scenarios.
- Bilingual Support: Trained on a curated dataset including both English and Chinese dialogues from the PIPPA dataset.
- Efficient Fine-tuning: Utilizes QLoRA with a LoRA Rank of 8, allowing for efficient adaptation.
Training Details
The model was trained for 1 epoch on 3,044 samples (1,522 English + 1,522 Chinese) with a maximum sequence length of 2048 tokens. The training process took approximately 45 minutes on an RTX 5080 16GB GPU, achieving a final evaluation loss of 1.9628.
Limitations
- Inherits biases from the PIPPA dataset and the base Qwen2.5-7B-Instruct model.
- Responses are typically constrained to around 30 Chinese characters (or equivalent), making it less suitable for long-form text generation.
- Character consistency may decrease for personas or scenarios outside the training data distribution.
When to Use This Model
This model is ideal for applications requiring short, character-driven conversational AI, particularly in bilingual contexts. It is well-suited for interactive storytelling, virtual companions, or any scenario where maintaining a specific persona in dialogue is crucial.