PriorTune

Code: https://github.com/yandex-research/PriorTune. The repository's bin/artifacts.py downloads and reads every file here.

Hyperparameter-configuration × dataset evaluation matrices. They are used to select small portfolios of hyperparameter configurations for tabular models by hill-climbing on synthetic datasets drawn from the TabICL v2 prior, then deploying the portfolios on real benchmarks.

For each of three model families, 4,000 random configurations were fit on each of:

• 4,000 synthetic datasets;
• the 51 TabArena datasets;
• 180 TabRepo datasets.

Every fit's validation and test predictions are included, together with the score matrices, the configurations, the synthetic datasets themselves and the selected portfolios.

Status (November 2026): upload in progress; the prediction folders fill shard by shard.

Models

name	what it is	configs
lgbm-prep	LightGBM; configs whose id is divisible by 4 also run TabPrep feature engineering	configs/lgbm-4000.jsonl
lgbm	plain LightGBM: the same 4,000 LightGBM hyperparameters, TabPrep off everywhere	configs/lgbm-4000.jsonl (ignore the tabprep fields)
mlp	MLP-Cos: the TabPack independent MLP with cosine feature embeddings	configs/mlp-4000.jsonl

The LightGBM search space is TabArena's (13 axes). Configs with id % 4 ≠ 0 are identical in lgbm and lgbm-prep.

Corpora

name	datasets	split
synthetic-tabicl-v2	4,000 datasets from the TabICL v2 prior (GraphSCM): 2,000 regression, 1,400 binary, 600 multiclass	iid 64 / 16 / 20, stratified
tabarena-51	the 51 TabArena datasets	outer repeat 0 / fold 0; validation = inner fold 0
tabrepo-180	180 TabRepo (AutoML 2024) datasets that do not overlap TabArena	test = OpenML repeat 0 / fold 0, validation = fold 1

Layout

configs/{lgbm,mlp}-4000.jsonl                    one config per line, line i = config_id i
scores/<corpus>/<model>.npz                      score matrices (datasets x configs)
predictions/<corpus>/<model>/part-NNNNN.parquet  raw validation / test predictions
labels/<corpus>.parquet                          y_val / y_test the predictions are scored against
datasets/tabicl-v2/part-NNNNN.parquet            the 4,000 synthetic datasets (features, target, split)
splits/{tabarena-51,tabrepo-180}.parquet         split indices + source ids of the real datasets
portfolios/<model>/{T,TE}_K{25,64,200}.json      selected portfolios

Score matrices (scores/…npz)

Each file is the exact array set the selection code consumes:

key	shape	description
validation, test	(datasets, 4000) float64	native metric
available	(datasets, 4000) bool	False where a fit failed
status_code	(datasets, 4000)	per-cell status
datasets	(datasets,)	dataset names
task_types	(datasets,)	regression / binclass / multiclass
metrics	(datasets,)	the native metric of each dataset

Native metrics are ROC-AUC (binary, higher is better), cross-entropy (multiclass, lower is better) and RMSE (regression, lower is better).

import numpy as np
from huggingface_hub import hf_hub_download
z = np.load(hf_hub_download('mrsergazinov/PriorTune', 'scores/synthetic-tabicl-v2/lgbm-prep.npz',
                            repo_type='dataset'))
val, test, ok = z['validation'], z['test'], z['available']

Predictions (predictions/…parquet)

One row per (dataset, split, block of 500 configs); shards are in dataset order and a dataset may span consecutive shards:

column	description
dataset_index, dataset	position in the corpus (= row of the score matrix) and dataset name
split	val or test
config_id_start, config_id_stop	the config ids of the block, [start, stop)
shape	[configs, rows], or [configs, rows, classes] for multiclass
dtype	numpy dtype of values (<f4)
values	raw bytes, row-major over shape; NaN where the fit failed

lgbm folders hold only the 1,000 TabPrep-off reruns (config ids 0, 4, …, 3996). Its other 3,000 configs are byte-identical to lgbm-prep (the same LightGBM fits, no TabPrep), so they are not stored twice. lgbm rows carry an explicit config_ids list (and config_id_start / config_id_stop = first id / last id + 1). To build the full lgbm array of a dataset, take lgbm-prep and overwrite the rows config_ids with the lgbm blocks. config_ids is null in every other model's rows.

Binary tasks hold the positive-class probability, multiclass tasks the class probabilities, regression tasks the prediction (on the standardised target for the synthetic corpus).

import numpy as np, pyarrow.parquet as pq
rows = pq.read_table(path, filters=[('dataset_index', '=', 7), ('split', '=', 'val')]).to_pylist()
pred = np.concatenate([np.frombuffer(r['values'], r['dtype']).reshape(r['shape'])
                       for r in sorted(rows, key=lambda r: r['config_id_start'])])   # (4000, n_val[, k])

Labels (labels/<corpus>.parquet)

Columns dataset_index, dataset, split (val, test), shape, dtype, values (raw bytes). These are the targets that the predictions and score matrices were scored against: int64 class ids, or float32 targets for regression.

Synthetic datasets (datasets/tabicl-v2/…parquet)

One row per dataset:

• dataset_index, dataset: the dataset's row in the synthetic score matrices, and its name;
• x_num, x_cat, y: .npy bytes, read with np.load(io.BytesIO(...));
• train, val, test: row indices;
• info: JSON with the generator seeds and shapes.

Real datasets (splits/…parquet)

Only split indices and provenance are included: OpenML task and dataset ids in info, and the source name. Fetch the data itself from its original source (TabArena / OpenML).

Portfolios (portfolios/…json)

T portfolios are selected for deploying the validation-best member (tuned). TE portfolios are selected for deploying a Caruana-style ensemble of the portfolio (tuned + ensembled). Each file has:

• config_ids;
• the greedy add_order;
• the full resolved configs;
• the selection protocol.

portfolios/tabpack/ holds the K = 64 portfolios that TabPack trains as a single 64-member pack; they are the same ids as portfolios/mlp/*_K64.json.

License and citation

To be added.