做了一个儿童学习助手开源 App，(可以理解为文本+摄像头输入, ChatGPT 响应， TTS/3D Avatar 展示)，请大佬们给点反馈机制的意见

https://github.com/see2023/see_me_now
当然你也可以当做一个简单的 3D Avatar 版本的 ChatGPT 移动端 App 。
路过的大佬帮忙给 AI 新手点意见，目前测试下来，用 gpt-3.5-turbo 无法根据视频的摘要信息(姿态、表情)返回真正有效的内容，自己写死的规则也有点傻，4.0 申请还没消息..
需求很简单，就是看着孩子写作业，发现他走神的时候，给与适当的提示。

sendMessage (605 tokens) {
  max_tokens: 1000,
  model: 'gpt-3.5-turbo',
  temperature: 0.8,
  top_p: 1,
  presence_penalty: 1,
  messages: [
    {
      role: 'system',
      content: 'You only need to watch the child in front of you and maintain a normal learning state. \n' +
        'You are given two arrays of input, which respectively record the sitting posture and exciting smile value of the child at the uniformly sampled time points in the last minute. \n' +
        'The value range of sitting posture is 0 or 1, where 1 represents sitting posture. \n' +
        'The value of exciting is 0 or 1, where 1 represents excitement; when in a long-term excited state, you need to let him calm down. \n' +
        'For example, if you receive {sit: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0], exciting: [0, 0, 0, 0, 0, 0, 0, 0, 1, 0]}, you need to remind the child to sit upright; \n' +
        'if the excitement is not too high, there is no need to remind. \n' +
        'If you receive {sit: [0, 1, 1, 1, 1, 1, 1, 1, 1, 1], exciting: [1 ,1 ,1 ,1 ,1 ,1 ,0 ,1 ,1 ,1]}, \n' +
        'you need to praise the child for sitting more and more upright but being too excited and need to tell the child to calm down. \n' +
        'Next minute you will receive two arrays and then give corresponding reminders based on these two sets of data. \n' +
        'The response content should be short and easy to understand and should not exceed twenty words; encouragement is the main content and interesting. Please reply in Chinese.'
    },
    {
      role: 'user',
      content: '{ sit: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0], exciting: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]}',
      name: undefined
    },
    { role: 'assistant', content: '宝宝，坐得挺好的，继续保持哦！', name: undefined },
    {
      role: 'user',
      content: '{ sit: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0], exciting: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]}',
      name: undefined
    },
    {
      role: 'assistant',
      content: '宝宝，坐姿有点不好哦，记得要坐直一些呀！',
      name: undefined
    },
    {
      role: 'user',
      content: '{ sit: [0, 0, 1, 1, 1, 1, 1, 1, 1, 1], exciting: [0, 0, 0, 0, 0, 0, 1, 1, 1, 1]}',
      name: undefined
    }
  ],
  stream: false
}

chatgpt response 宝宝，坐得越来越好了呢！但现在有点过于兴奋了，让我们一起冷静下来吧！

难道应该用强化学习吗？好像也没有合适的奖励函数和训练环境