Datasets:

mdsajjadullah
/

ml-failures-dataset

license: mit
language:
  - en
pretty_name: ML Model Failures Dataset
size_categories:
  - n<1K
task_categories:
  - tabular-classification
tags:
  - machine-learning
  - debugging
  - model-failures
  - overfitting
  - underfitting
  - adversarial
  - bias
  - class-imbalance
  - gradient-problems
  - model-drift
  - out-of-distribution
  - mlops
  - education

ML Model Failures Dataset

What Is This

This dataset contains 900 annotated machine learning failure records across 9 failure types. Every record documents what went wrong, why it went wrong, and how to fix it.

Load It In Python

from datasets import load_dataset

ds = load_dataset("YOUR_HF_USERNAME/ml-failures-dataset") df = ds['train'].to_pandas()

print(df['failure_type'].value_counts()) print(df['severity'].value_counts())

The 9 Failure Types

Overfitting — High severity
Underfitting — High severity
Adversarial Attack — Critical severity
Bias — Critical severity
Class Imbalance — High severity
Gradient Exploding — High severity
Gradient Vanishing — High severity
Model Drift — High severity
Out-of-Distribution — Critical severity

All 14 Columns

scenario_id — unique ID like OVF_0001 failure_type — which failure category feature_values — input numbers as a list true_label — correct answer 0 or 1 train_accuracy — accuracy on training data val_accuracy — accuracy on test data precision — weighted precision score recall — weighted recall score f1_score — weighted F1 score error_description — what went wrong root_cause — why it failed fix_strategy — how to fix it severity — Critical, High, or Medium extra_info — extra diagnostic info

Average Metrics By Failure Type

Overfitting train 1.000 val 0.700 f1 0.700 Underfitting train 0.720 val 0.710 f1 0.710 Adversarial Attack train 0.880 val 0.650 f1 0.640 Bias train 0.820 val 0.780 f1 0.760 Class Imbalance train 0.960 val 0.950 f1 0.910 Gradient Exploding train 0.500 val 0.500 f1 0.490 Gradient Vanishing train 0.500 val 0.500 f1 0.490 Model Drift train 0.880 val 0.600 f1 0.590 Out-of-Distribution train 0.950 val 0.480 f1 0.470

How Data Was Created

All data is synthetic and reproducible using scikit-learn. Run ML_Failures_Dataset.ipynb in Google Colab to regenerate everything.

License

MIT — free for research, education, and commercial use.

Citation

@dataset{ml_failures_dataset_2026, title = {ML Model Failures Dataset}, year = {2026}, version = {1.0.0}, license = {MIT} }