ShareGPT is a dataset of user-provided ChatGPT conversations used to fine-tune conversational language models, including Vicuna, for virtual-assistant tasks. Its practical value comes from exposing models to real dialogue rather than relying only on general foundational-model training. section 10.3.13
ShareGPT supports several applied goals:
- Fine-tune open-source foundational models for domain-relevant conversational behavior. section 10.3.13
- Build virtual assistants with accuracy approaching larger models while using fewer parameters. section 10.3.13
- Support evaluation alongside chatbot arenas that compare competing conversational systems. section 10.3.13
Vicuna fine-tuning
Vicuna was trained on 90,000 ShareGPT conversations. In 2023, the dataset contained almost half a million ChatGPT conversations, illustrating its growth as an open conversational-data resource. section 10.3.13
