Bài kiểm tra có liên kết hướng dẫn · Trung bình Cấp độ

Đánh giá bài kiểm tra hội thoại nhiều lượt

Đánh giá cuộc đối thoại hoàn chỉnh về bối cảnh, thành công của nhiệm vụ, khả năng phục hồi và an toàn, đồng thời kiểm tra các mẫu điểm chuẩn và giới hạn đánh giá tự động.

Đường dẫn hướng dẫn liên quanĐánh giá cuộc trò chuyện nhiều lượt
Câu hỏi 1 của 8

Why evaluate a conversation across multiple turns?

Xem trước mọi câu hỏi8 câu hỏi

Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.

  1. Why evaluate a conversation across multiple turns?
    • Single-turn scores always measure long-term memory
    • Multi-turn tests remove all uncertainty
    • Only the last message matters
    • A later response can fail to use earlier context or complete the overall goal
  2. Which dimension may matter in a multi-turn support dialogue?
    • Only the first token latency
    • Whether the assistant retained earlier facts and resolved the user’s issue
    • Only the number of total turns
    • The chatbot’s training loss
  3. What was the scope of MT-Bench-101 cited in the guide?
    • A million production conversations from every industry
    • Only single-turn math prompts
    • 1,388 dialogues across 13 tasks, comprising 4,208 turns
    • One conversation judged by one person
  4. Does a benchmark score guarantee performance in every deployed conversation?
    • No; results apply to the benchmark tasks and sample
    • Yes, if the benchmark has more than 1,000 turns
    • Yes, if an automated judge was used
    • No, because benchmarks provide no information
  5. Why validate an automated conversation judge against human review?
    • To guarantee every human judge agrees
    • To remove the need for test cases
    • To make dialogue length irrelevant
    • To check agreement, rubric quality, and possible judge bias
  6. What should a dialogue test do when a user corrects an earlier instruction?
    • Ignore the correction and score only the first response
    • End the test immediately
    • Check whether the assistant incorporates the correction in later turns
    • Treat the correction as a tool result
  7. Which cases can expose dialogue-level weaknesses?
    • Only short greetings
    • Long histories, interruptions, tool failures, and safety-sensitive turns
    • Only repeated identical prompts
    • A model card without conversations
  8. How should a team report multi-turn evaluation?
    • Report sample selection, scoring rules, and failure categories
    • Publish only a single average score
    • Hide judge disagreement
    • Report only answer length