ampligraph.datasets.load_fb15k_237(check_md5hash=False, clean_unseen=True, add_reciprocal_rels=False)

Load the FB15k-237 dataset

FB15k-237 is a reduced version of FB15K. It was first proposed by [TCP+15].

The FB15k-237 dataset is loaded from file if it exists at the AMPLIGRAPH_DATA_HOME location. If AMPLIGRAPH_DATA_HOME is not set the the default ~/ampligraph_datasets is checked.

If the dataset is not found at either location it is downloaded and placed in AMPLIGRAPH_DATA_HOME or ~/ampligraph_datasets.

The dataset is divided in three splits:

  • train
  • valid
  • test
Dataset Train Valid Test Entities Relations
FB15K-237 272,115 17,535 20,466 14,541 237


FB15K-237’s validation set contains 8 unseen entities over 9 triples. The test set has 29 unseen entities, distributed over 28 triples.

  • check_md5hash (boolean) – If True check the md5hash of the files. Defaults to False.
  • clean_unseen (bool) – If True, filters triples in validation and test sets that include entities not present in the training set.
  • add_reciprocal_rels (bool) – Flag which specifies whether to add reciprocal relations. For every <s, p, o> in the dataset this creates a corresponding triple with reciprocal relation <o, p_reciprocal, s>. (default: False).

splits – The dataset splits: {‘train’: train, ‘valid’: valid, ‘test’: test}. Each split is an ndarray of shape [n, 3].

Return type:



>>> from ampligraph.datasets import load_fb15k_237
>>> X = load_fb15k_237()
>>> X["train"][2]
array(['/m/07s9rl0', '/media_common/netflix_genre/titles', '/m/0170z3'],