TigerBot is a suite of multilingual, multi-task Large Language Models (LLMs) designed to serve as a foundation for custom LLM applications, particularly targeting Chinese innovation. It offers a range of model sizes and versions, built upon Llama-2 and Bloom architectures, with continuous updates and performance improvements.

How It Works

TigerBot models are trained on extensive datasets, including curated Chinese and English text from books, the internet, and encyclopedias. The training process incorporates techniques like holistic training, GQA, flash-attn, and RoPE for efficiency and performance. Alignment is achieved through methods such as grouped SFT, rejection sampling, and DPO. Recent versions support extended context lengths up to 100K tokens using YaRN for extrapolation.

Quick Start & Requirements

• Installation: conda create --name tigerbot python=3.8, conda activate tigerbot, conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia, git clone https://github.com/TigerResearch/TigerBot, cd TigerBot, pip install -r requirements.txt.
• Prerequisites: Python 3.8+, PyTorch with CUDA 11.7. GPU is required for inference and training (e.g., A100 40GB for 7B models, multiple A100s for larger models).
• Demo: streamlit run apps/web_demo.py -- --model_path tigerbot-13b-chat
• Docs: TigerBot API

Highlighted Details

• Supports up to 100K context length with YaRN.
• Offers function calling capabilities.
• Includes RAG features (search and document augmentation).
• Claims SOTA performance on various Chinese and English benchmarks.
• Provides open-source pre-training and fine-tuning datasets.

Maintenance & Community

The project is actively maintained with frequent updates and new model releases. Community engagement is encouraged via discussion groups.

Licensing & Compatibility

Models are released under a free commercial license. The project aims for compatibility with frameworks like LangChain and supports OpenAI-compatible APIs.

Limitations & Caveats

The project acknowledges potential issues such as model hallucination, misleading content, or discriminatory output. Users are advised to use generated content cautiously and are responsible for any harmful content disseminated.