rank

name	rank
description	Rank models per pred_len/seed for a dataset, producing MSE and MAE leaderboard CSVs. Use when the user wants leaderboards, rankings, or to compare model performance across prediction horizons.

When to use

After experiments have produced performance.csv files under work_dirs/. Ask the user for the dataset name if not provided.

Command

uv run python tool/rank_models.py --dataset <DATASET>

Writes model_rankings_mse.csv, model_rankings_mae.csv, and model_rankings_long.csv to work_dirs/<DATASET>/. Rankings are computed per (pred_len, seed) group; rank 1 = lowest metric value.

For a fair leaderboard (TFB-style), add --null-threshold 0.3 (exclude models missing on too many cells), --aggregate mean|median|max (collapse duplicates), and/or --fill-nan-with-mean.

Reference

Full flags, input requirements, and output formats: docs/en/rank-models.md (or tool/rank_models.py --help).