Dataset is linked from the mlx-community Space instead

Bench: don't report speed when trials disagree

Use perplexity from mlxbench-2 only (BOS in every window)

Update MLX Model Explorer

Reload stored events every 15 minutes (and after boot) so stats and rankings never lose data

Record one summary per visit instead of every click; hourly writes to data/incoming; monthly compaction

Measured quality: perplexity per quantization, compared within base model and bit width

Keep filters open on wide layouts

Deploy MLX Model Explorer (private test)

Duplicate from autotrain-projects/autotrain-advanced