This page hosts no data. PaMIR does not redistribute the datasets. It is a Python package (pip install "pamir[data]") that fetches each dataset from its original source and harmonizes it locally. The apache-2.0 tag covers the package code only; each dataset stays under its own license.

PaMIR

Public Arrival-ordered Measurement for Inference in Risk

An open benchmark for credit-default prediction when labels are scarce and arrive late: 19 public credit-default datasets (1.24M loans, firms and card accounts from nine countries, default rates 3%–41%), rebuilt from pinned source snapshots by one leakage-audited recipe and never redistributed; to our knowledge it is the one of its kind as of today. Every model is scored under two evaluation protocols: a label-delayed stream, in which each application is scored on arrival by a model trained only on outcomes that have matured, with AUC by label budget, and a repeated i.i.d. split for comparison with other tabular benchmarks. Fleet means are withheld unless every dataset is scored. A leakage-controlled synthetic-data harness tests generated training rows. This card hosts no data.

Quick use

# pip install "pamir[data]"
# Fetches gmsc from its original source and harmonizes it locally on first use.
from pamir import load_dataset, evaluate, evaluate_iid, gbdt_fit, gbdt_baseline
X, y, meta = load_dataset("gmsc")
results = evaluate(gbdt_fit)                          # streaming, reference setting

results_iid = evaluate_iid(gbdt_baseline, n_seeds=5)  # conventional i.i.d.

What makes this benchmark different

General tabular benchmarks contain credit tasks but evaluate them on random, temporal or grouped splits. A credit model is trained on outcomes that mature months or years after origination and starts with no labels; a temporal split orders the data by time but can still train on outcomes that matured after its cut-off. PaMIR's streaming protocol scores each application on arrival, with a model trained only on outcomes that had matured by then, and reports how the AUC grows with the number of labels. Rows are replayed in a fixed random order (most sources carry no usable dates), so it measures learning under scarce, delayed labels, not robustness to temporal shift.

Benchmark	Datasets	Domain	Evaluation
OpenML-CC18	72	General	i.i.d. cross-validation
TabArena	51	General	i.i.d. cross-validation
MultiTab	196	General	i.i.d. splits
TabReD	8	Industrial	time-based splits
BeyondArena	142	General	i.i.d., temporal and grouped splits
Lessmann et al. (2015)	8 (4 public)	Credit	i.i.d. splits
PaMIR	19	Credit default	i.i.d. and label-delayed streaming

Datasets

id	name	rows	DR	geography	product	license
bankruptcy	Taiwanese Bankruptcy Prediction	6,819	3.2%	Taiwan	Corporate	Data files © Original Authors (Kaggle)
bondora	Bondora P2P Lending	266,482	40.9%	Estonia, Finland, Spain	P2P consumer	CC0 1.0 (Kaggle reuploader)
conorsully	Conor Sully Credit Score	1,000	28.4%	Synthetic / educational	Consumer	CC0 1.0
dish	Automobile Loan Default	121,856	8.1%	Unspecified	Vehicle finance	CC0 1.0 (Kaggle reuploader)
gastonstat	Gaston Sanchez Credit Scoring	4,454	28.1%	Unknown	Consumer	No license (GitHub, no LICENSE file)
gmsc	Give Me Some Credit	150,000	6.7%	USA	Consumer revolving + installment	Unknown (Kaggle reuploader)
laotse	Laotse Credit Risk	32,581	21.8%	Unspecified	Consumer	CC0 1.0 (Kaggle reuploader)
lc_clean	Lending Club (cleaned, 2007–2014)	150,000	20.2%	USA	P2P consumer	No license (HuggingFace repo, none stated)
lc_my	Lending Club (Malaysian variant)	100,000	22.6%	Unspecified	Consumer	Unknown (Kaggle reuploader)
lc_small	Lending Club (small, 9,578 loans)	9,578	16.0%	USA	Consumer	ODbL (Open Database License)
lt_vehicle	L&T Vehicle Loan Default	233,154	21.7%	India	Vehicle finance	Other (specified in description)
pakdd	PAKDD 2010 Credit Data	50,000	26.1%	Brazil	Consumer	No license (competition archive)
poland_1yr	Polish Companies Bankruptcy (1-year horizon)	7,027	3.9%	Poland	Corporate	CC-BY-4.0 (UCI)
poland_3yr	Polish Companies Bankruptcy (3-year horizon)	10,503	4.7%	Poland	Corporate	CC-BY-4.0 (UCI)
poland_5yr	Polish Companies Bankruptcy (5-year horizon)	5,910	6.9%	Poland	Corporate	CC-BY-4.0 (UCI)
prosper	Prosper Marketplace Loans	55,084	30.9%	USA	P2P consumer	CC0 1.0 (Kaggle reuploader)
sba	U.S. SBA Loan Defaults	2,102	32.6%	USA	Small business	CC0 1.0 (Kaggle reuploader)
south_german	South German Credit (corrected)	1,000	30.0%	Germany	Consumer	CC-BY-4.0 (UCI)
taiwan	Taiwan Credit Card Default	30,000	22.1%	Taiwan	Credit card	CC0 1.0 (Kaggle “UCI ML” reuploader)

Total: 1,237,550 rows, 281,766 defaults across 19 datasets.

Streaming protocol (summary)

1. Applications arrive in stored order (a fixed random permutation).
2. The outcome of the application at position t is revealed at t + lag (default 1000, capped at 20% of the stream).
3. The model is refitted after every k newly resolved defaults (default 10).
4. Each application is scored once, on arrival, by the latest model; scores are final.
5. Metric: ROC AUC over all scored applications, and AUC by label budget; the fleet mean is withheld unless every dataset is scored.

Full protocol specification and evaluation code: pip install pamir (paper, documentation, code).

A repeated i.i.d. split protocol (evaluate_iid) is also provided for comparison with other tabular benchmarks.

Data provenance

PaMIR ships no data. pamir.download(id) fetches each dataset from its original source and harmonizes it with a reproducible recipe: derive the binary target, drop id/date/constant/surrogate columns, drop post-outcome columns by the elicitation's semantic cut (day_zero_available = false) plus per-dataset extra drops, and shuffle rows with a fixed seed. The source snapshot is pinned, and the rebuilt table must match its data contract in expected.json exactly (columns, counts, SHA-256 of the raw file and of the target vector) or it is refused. Per-dataset recipe: pamir.dataset_info(id) in the Python package.

License

The PaMIR package code is Apache 2.0. PaMIR does not redistribute the datasets — each is fetched from its original source at the user's request and remains under its own license and terms (listed in the table above and in pamir.dataset_info(id)). Users are responsible for complying with each source's terms and citing its original authors.

Citation

@misc{liashkov2026pamir,
  title         = {PaMIR: Open Benchmark of Public Credit-Default Datasets},
  author        = {Liashkov, Mikhail and Varshavskiy, Ilyas and Khalilbekov, Shuhratjon and
                   Azimi, Azizjon and Boboeva, Bonu},
  year          = {2026},
  eprint        = {2610.03259},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2610.03259},
}

@software{pamir2026,
  title   = {PaMIR: Public Arrival-ordered Measurement for Inference in Risk},
  author  = {Liashkov, Mikhail and Varshavskiy, Ilyas and Boboeva, Bonu and
             Khalilbekov, Shuhrat and Azimi, Azizjon},
  year    = {2026},
  version = {0.4.0},
  url     = {https://github.com/zypl-ai/pamir},
}