LLM Zoo provides data, models, and evaluation benchmarks for large language models, aiming to democratize ChatGPT-like capabilities across languages. It targets researchers and developers seeking to build, evaluate, and deploy multilingual instruction-following and conversational LLMs. The project offers open-source models and training code, enabling replication and customization.

How It Works

LLM Zoo's models, Phoenix and Chimera, are trained on a combination of instruction data (self-instructed/translated and user-centered) and user-shared conversation data. This dual-data approach aims to imbue models with both instruction adherence and conversational fluency, addressing limitations of models trained on only one data type. Phoenix is a multilingual model based on BLOOMZ, while Chimera is based on LLaMA and focuses on Latin and Cyrillic languages.

Quick Start & Requirements

• Install via pip install -r requirements.txt.
• CLI inference: python -m llmzoo.deploy.cli --model-path <model_name_or_path>.
• Models are available on Hugging Face (e.g., FreedomIntelligence/phoenix-inst-chat-7b).
• Chimera requires applying delta weights to base LLaMA models.
• Deployment involves launching controller, model worker, and Gradio web server.
• See official documentation for detailed deployment and training instructions.

Highlighted Details

• Phoenix-inst-chat-7b achieves 85.2% of ChatGPT's performance in Chinese via GPT-4 evaluation and leads open-source Chinese LLMs.
• Offers int8 and int4 quantization for reduced GPU memory consumption (e.g., ~7GB for Phoenix).
• Provides delta weights for LLaMA-based Chimera models due to LLaMA's license restrictions.
• Includes evaluation benchmarks and training scripts for replicating models.

Maintenance & Community

The project is primarily contributed by researchers from The Chinese University of Hong Kong, Shenzhen. Contributions are welcomed via the GitHub repository.

Licensing & Compatibility

Models are released under a license that permits use and modification. However, Chimera models require original LLaMA weights, which have their own licensing terms.

Limitations & Caveats

The project acknowledges limitations common to LLMs, including lack of common sense, limited knowledge domains, potential biases inherited from training data, and difficulties in understanding emotions or nuanced context. Benchmarking is noted as a challenging task.