kandpal-2023-bin-trimming-500-min
IN premise — summaries/2026/08/24/kandpal-2023-long-tail-knowledge-s2-identifying-relevant-pre-training-data.md
Created 2026-08-25T02:58:06+00:00
Binned accuracy plots in Kandpal et al. (2023) trim bins containing fewer than 500 examples.
Summary
When Kandpal and colleagues visualize how model accuracy varies across different groups of inputs, they exclude any group with fewer than 500 samples. This means the plots only reflect groups large enough to be statistically meaningful, so readers aren't led to draw conclusions from tiny, noisy subsets of data.