Closed tttonytan closed 5 months ago
@hsaest good day. I find that the prompt in solo planning (planner_agent_prompt.format(text=ii['reference_information'], query=ii['query'])) is quite lengthy, usually around 10K even up tp 16K.
Therefore, can I say that, this is the primary reason why in the experiment results, for open-llm, only Mixtral-8*7B-MoE is used for testing. Since most other open LLMs cannot handle such long input context (except "lmsys/vicuna-7b-v1.5-16k"
"mistralai/Mistral-7B-Instruct-v0.2" (32K). Am I right ?
(for close-llms. open-35-turbo, 16K context length is accepted)
@yananchen1989 True. We only choose these open-sourced models that support more than 8K context.
Hi,
Thank you for your interest in our work.
We offer a training set to encourage further follow-up research based on agent learning (e.g., in-context learning, reinforcement learning) within the TravelPlanner. We've assessed baseline performance of current agent frameworks in a zero-shot setting, without training LLMs.
Feel free to contact us if you have further questions.