KULLM (구름) is a family of Korean-centric Large Language Models developed by Korea University's NLP & AI Lab and HIAI Research Institute. It aims to provide high-quality Korean language understanding and generation capabilities, building upon existing models like Polyglot-ko and SOLAR.

How It Works

KULLM models are instruction-tuned on Korean datasets. KULLM3, the latest iteration, is based on upstage/SOLAR-10.7B-v1.0 and was trained using 8x A100 GPUs. The training incorporates a specific system prompt designed to guide the AI's persona and behavior, emphasizing helpfulness, accuracy, and adherence to ethical guidelines.

Quick Start & Requirements

• Install with: pip install torch transformers==4.38.2 accelerate
• Requires CUDA-enabled GPU.
• Tested with transformers>=4.40.0.

Highlighted Details

• KULLM3 is available in quantized (AWQ 4-bit) and full precision (fp16) formats.
• Evaluation uses G-Eval with GPT-4-Turbo, employing a translated version of the yizhongw/self-instruct dataset.
• English prompts were found to yield more accurate evaluation results.

Maintenance & Community

The project is actively maintained by Korea University's NLP & AI Lab. Citation details for previous versions and the current KULLM3 are provided.

Licensing & Compatibility

Licensed under Apache-2.0, allowing for commercial use and integration into closed-source projects.

Limitations & Caveats

The model may exhibit hallucination and repetition depending on the decoding strategy. Generated outputs can be inaccurate or harmful. Performance may be lower on benchmarks that do not utilize the fixed system prompt used during training.