HPLT × NLPL Winter School · Day 2 · 11 February 2025 · Scroll to read
About these notes
Lecture material on evaluation leakage, benchmark contamination, contamination-resistant tests, and the methodological traps that can make language-model generalization look stronger than it is.
- Training/test leakage and benchmark contamination
- Dynamic and contamination-resistant evaluations
- Reliable measurement of model generalization
-
Page 1 of 21 -
Page 2 of 21 -
Page 3 of 21 -
Page 4 of 21 -
Page 5 of 21 -
Page 6 of 21 -
Page 7 of 21 -
Page 8 of 21 -
Page 9 of 21 -
Page 10 of 21 -
Page 11 of 21 -
Page 12 of 21 -
Page 13 of 21 -
Page 14 of 21 -
Page 15 of 21 -
Page 16 of 21 -
Page 17 of 21 -
Page 18 of 21 -
Page 19 of 21 -
Page 20 of 21 -
Page 21 of 21