Koala
lastmod 2023-04-10created 2023-04-101 edits
Dialogue model for academic research (LLaMA-based chatbot)
Features
- Koala-13B model, which, like Alpaca, trains LLaMA on dialogue and instruction sets
- Produces results preferred over Alpaca on a variety of queries, and can generate responses on par with ChatGPT in at least half of the cases
- Supervised Fine-Tuning on open-source dialogue data collected from the web
- 60K dialogues from ShareGPT
- 87K question-answer examples from HC3
- Open Instruction Generalist (OIG)
- 52K dataset released by Stanford Alpaca
- Anthropic HH (160K)
- OpenAI WebGPT (20K))
- OpenAI Summarization (93K)