When N-gram Dedup Fails: How Much Paraphrasing Hides a Benchmark Leak
A controlled leak-detection test: word n-gram overlap catches every copy-pasted benchmark duplicate but only 25-47% of synonym-paraphrased ones.
-
-
Reproducing Double Descent: Why 300 Features Beat 39 on the Same 40 Data Points
A minimal least-squares experiment reproduces double descent: test error peaks near the interpolation threshold, then drops 28x as parameters keep growing.
-
How Much Real Data Stops Model Collapse? A Minimal Recursive-Training Simulation
A 500-run simulation shows recursive synthetic training can shrink variance by up to 81% in 40 generations; mixing in 10% real data cuts that to under 2%.
