20,000 training samples for the experimental theme, generated by DeepSeek.
Each sample contains source abstracts from a scientific cluster and target outputs (abstract summary, short summary, title, overview) for training the BSG CyLlama cluster description generator.