指令调整测验

当您参加测验时，就会出现答案和解释。

1. 如果没有指令调整，原始基础模型如何响应“法国的首都是哪里？”的提示？
   • 总是拒绝回答
   • 它返回一个错误
   • 它可能会继续提出更多问题或相关文本，而不是回答
   • 每次都完美回答
2. 指令调整最好描述为：
   • 从人类反馈中强化学习
   • 对原始网络文本进行预训练
   • 调节温度参数
   • 对指令和响应对进行监督微调
3. How does supervised instruction tuning differ from reward-model-based RLHF in the published InstructGPT recipe?
   • It is exactly the same optimization stage
   • It trains against reference responses, while the later preference stages use comparisons and a reward model
   • It requires no training examples
   • It only changes a sampling temperature
4. Why do instruction-tuning researchers evaluate models on held-out tasks and instructions?
   • To assess generalization beyond the examples used for training
   • To prove that every unseen task will succeed
   • To avoid checking any generated responses
   • To guarantee the same accuracy for every task mixture
5. 现代聊天助理的典型培训阶段顺序是什么？
   • RLHF，然后预训练，然后指令调优
   • 指令调优，然后预训练
   • 只是预训练，没有其他
   • 预训练，然后指令调整，然后可选的 RLHF
6. What does LIMA's curated-data experiment support as a research conclusion?
   • Exactly 1,000 examples are enough for every model and task
   • More data can never help any instruction-tuned model
   • Carefully selected small datasets can be effective in an evaluated configuration
   • Only synthetic examples can train an assistant
7. Which training stage did the original LIMA experiment omit while using its 1,000 curated prompt-response pairs?
   • Reinforcement learning and human preference modeling
   • Supervised response loss
   • A pretrained base model
   • Prompt-response examples
8. A tuned model follows the requested answer format. What does that behavior alone fail to establish?
   • That an output format was requested
   • That every factual statement in the answer is correct
   • That the model produced a response
   • That the response can be inspected