Bài kiểm tra có liên kết hướng dẫn · Trung bình Cấp độ
Bài kiểm tra ô nhiễm điểm chuẩn
Kiểm tra xem các mục điểm chuẩn rò rỉ vào dữ liệu đào tạo như thế nào, các phương pháp phát hiện như chồng chéo n-gram và Kiểm tra Min-K% cũng như các biện pháp giảm nhẹ như canary và phân chia thời gian.
Câu hỏi 1 của 8
What is benchmark contamination?
Xem trước mọi câu hỏi8 câu hỏi
Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.
- What is benchmark contamination?
- When a benchmark contains incorrect answer keys
- When a model is tested on too few questions
- When benchmark test items or their answers appear in a model's training data
- When two benchmarks measure the same skill
- What is the purpose of the canary string included in BIG-bench tasks?
- To let data teams filter those documents from training data and test whether a model saw them
- To encrypt the answers
- To mark which tasks are the hardest
- To identify which model generated each answer
- What did the GSM1k study suggest?
- All models scored higher on the new problems
- GSM8K has no errors
- Contamination cannot be measured without training data
- Some models scored noticeably lower on fresh, matched problems than on GSM8K
- What overlap analysis did the GPT-3 paper run?
- Embedding similarity on full documents
- 13-gram overlap between benchmarks and training data
- Human review of every training document
- Min-K% Prob on each benchmark
- What is the core idea of Min-K% Prob?
- Removing the least common k percent of words from training data
- Training a second model to classify contaminated text
- Text the model has seen tends to have few very low-probability tokens
- Measuring the length of the model's answers
- How does a time-based split, as in LiveCodeBench, help detect contamination?
- It compares performance on problems published before and after the model's training cutoff
- It measures how long the model takes to answer
- It splits the model into older and newer layers
- It only counts problems answered within a time limit
- Why does finding no exact overlap not prove a benchmark is clean?
- Exact matching is too slow to run
- Benchmarks change their answers over time
- Models never memorize exact text
- Paraphrased or translated copies can slip past string matching
- In n-gram decontamination, what happens if n is chosen too small?
- Near-copies with small edits are missed
- Many false positives from common phrases are flagged
- The filter stops working entirely
- Only answers, not questions, are matched