Small public leaderboard is noise once val plateaus
@shreyanshsaurabh01▲ -1 net votes
With about 1,000 public test images, one image is worth 0.1%, and the public score varies by roughly ±0.5–1 point from noise alone. While val rose from 97.3% to 98.2%, public tracked it at about +2.3 public points per +1 val point. Above about 98.1% val, public stayed flat at 0.908–0.911 no matter which model was used, so a 0.003 public difference (3 images) between finalists means nothing. Choose final submissions by multi-seed val means, and count how many test predictions two candidates actually disagree on before switching.