Traditional Chinese QLoRa Llama2 13b Merged Model

This model, developed by weiren119, is a 13 billion parameter Llama 2-based language model specifically fine-tuned for Traditional Chinese instruction-following tasks. It utilizes the QLoRA (Quantized Low-Rank Adaptation) method, enabling efficient fine-tuning on consumer-grade hardware (e.g., an RTX 3090 with 24GB VRAM).

Key Capabilities

• Traditional Chinese Proficiency: Optimized for understanding and generating text in Traditional Chinese.
• Instruction Following: Fine-tuned on a translated version of the Stanford Alpaca 52k dataset, enhancing its ability to follow instructions.
• Efficient Deployment: Merged QLoRA adapters allow for direct use of the fine-tuned model.
• Llama 2 Foundation: Benefits from the robust architecture and pre-training of the Llama 2 chat model.

Good For

• Applications requiring a strong understanding and generation of Traditional Chinese.
• Building chatbots or assistants that interact in Traditional Chinese.
• Research and development in Traditional Chinese NLP with a Llama 2 base.
• Scenarios where efficient, quantized models are preferred for deployment or resource constraints.

This model was trained using bitsandbytes 4-bit quantization with nf4 type and bfloat16 compute dtype, and PEFT 0.4.0. Resources for the 7B version and an online Colab demo are also available.