Question about multi-round conversations.

[Li-Qingyun]

Hi, I find your work and technical readmap truly impressive. I have a question: since the florence-2 uses encoder-decoder architecture for language model, how does the multi-round conversation in the demo video implement? Is the incontext conversations history included in the encoder input?

By the way, can you review the model request of me? Thanks a lot!

[KeluYao]

Hi, I find your work and technical readmap truly impressive. I have a question: since the florence-2 uses encoder-decoder architecture for language model, how does the multi-round conversation in the demo video implement? Is the incontext conversations history included in the encoder input?

By the way, can you review the model request of me? Thanks a lot!

Dr. li ,
Thank you for your interest. The multi-turn conversation shown in the demo video is implemented within the demonstration system. Specifically, each interaction is processed as an independent single-turn dialogue, meaning that previous conversation history is not used as contextual input for the encoder.
Additionally, your request for model weights has been approved.

[Li-Qingyun]

Hi, thanks for your response and review.
I think i misunderstand between the multi-round and multi-instruction.

By the way, thanks for thre approvement of TianHuiLab/Falcon-Multi-Instruction-Large, I'm also insterested in the other two model, which have not been approved.