Giters

KwaiKEG / KwaiAgents

A generalized information-seeking agent system with Large Language Models (LLMs).

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Questions enquiry for KwaiAgentSys and MAT model

superobk opened this issue · comments

commented

Hello, here are some questions and want to understand. Please kindly assist.
1.
May I know how to use kagentsys in conversation mode? Currently with default parameters in README, it can only have one question for prompt input with 5 iterations maximum.
2.
For MAT model of Baichuan2 published, the base training model is [Baichuan2-13B-Chat] or [Baichuan2-13B-Base]? If MAT finetuning is based on [Baichuan2-13B-Base], could we have some way to finetune well in [Baichuan2-13B-Chat]?
3.
Did MAT models tested in other Agents framework, such as LangChain or AutoGPT? If yes, how is the performance.

Thank you in advance!

commented

Hi, thanks for your questions.

  1. You can directly use AgentService class in kwaiagents/agent_start.py and pass history to activate conversation mode. We will give a tutorial soon in README.md
  2. The Baichuan2-13B-Chat is used as the base model.
  3. Yes, you can check the benchmark results both in README and our papers. In Langchain, actually they use ReAct prompt templates.
commented

Hi, thanks for your questions.

  1. You can directly use AgentService class in kwaiagents/agent_start.py and pass history to activate conversation mode. We will give a tutorial soon in README.md
  2. The Baichuan2-13B-Chat is used as the base model.
  3. Yes, you can check the benchmark results both in README and our papers. In Langchain, actually they use ReAct prompt templates.

Thanks a lot, I will close the case.