sharpbai/openchat
OpenChat is a 13 billion parameter language model developed by OpenChat, fine-tuned on a high-quality dataset of approximately 6,000 GPT-4 multi-round conversations. Based on the LLaMA architecture, this model is optimized for conversational AI, achieving 105.7% of ChatGPT's score on Vicuna GPT-4 evaluation. It is particularly effective for general-purpose chat applications and interactive dialogue systems.
Loading preview...
OpenChat: Less is More for Open-source Models
OpenChat is a series of open-source language models, with this specific variant being a 13 billion parameter model based on LLaMA. It is uniquely fine-tuned on a compact yet high-quality dataset of approximately 6,000 GPT-4 multi-round conversations, demonstrating that effective performance can be achieved with limited data.
Key Capabilities & Performance
- High Conversational Performance: Achieves 105.7% of ChatGPT's score on the Vicuna GPT-4 evaluation, indicating strong capabilities in understanding and generating human-like dialogue.
- Efficient Training: Utilizes only ~6K GPT-4 conversations for fine-tuning, showcasing data efficiency.
- AlpacaEval Win-rate: Boasts an 80.9% win-rate on AlpacaEval, further validating its conversational prowess.
- Context Length: The base OpenChat model supports a 2048 token context length, with an extended version (OpenChat-8192) supporting 8192 tokens.
Use Cases
OpenChat is well-suited for applications requiring robust conversational AI, including:
- General-purpose chatbots: Excelling in multi-round dialogue scenarios.
- Interactive assistants: Providing engaging and coherent responses.
Technical Details
The model uses a specific conversation template involving concatenation of tokens and an end-of-turn token <|end_of_turn|>. The project also provides an inference server compatible with the "ChatCompletions" API and a web UI for enhanced user experience.