TDGBench

TDGBench provides immutable graph artifacts, Gold Standards and ten fixed train/validation/test splits for reproducible evaluation of text-driven knowledge graphs with graph neural networks.

Contents

biomed/
  graphs/GT2KG_tdg/{default,raw,no_smnt,no_freq_filter,no_freq_filter_no_smnt}.json
  graphs/kggen/{default,raw,no_smnt,no_freq_filter,no_freq_filter_no_smnt}.json
  corpus/medmentions_mapped.json
  gold_standard/gold_standard.xlsx
  splits/split_<seed>.json
  manifest.json
dbpedia/
  graphs/GT2KG_tdg/default.json
  graphs/cg/default.json
  corpus/dbpedia_174_abstracts.json
  gold_standard/gold_standard.xlsx
  splits/split_<seed>.json
  manifest.json
manifest.json

The predefined split seeds are 42, 123, 456, 789, 2024, 2025, 2026, 2027, 2028 and 2029. Every manifest records the byte size and SHA-256 digest of each benchmark asset.

Biomed preprocessing options

• raw: no preprocessing;
• no_smnt: without semantic-type integration;
• no_freq_filter: without relation-frequency filtering;
• no_freq_filter_no_smnt: both ablations together;
• default: the fully processed published graph.

DBpedia has no preprocessing variants.

Corpora

• biomed/corpus/medmentions_mapped.json: the mapped MedMentions corpus used to construct the biomedical TDGs;
• dbpedia/corpus/dbpedia_174_abstracts.json: the DBpedia-174 abstract corpus used to construct the DBpedia TDG.

The corpus assets are declared in their domain manifest and protected by byte-size and SHA-256 checks.

Use with tdg-bench

The companion Python library is maintained at OthmaneKabal/tdg-bench. It loads these artifacts, validates their checksums, applies the fixed splits and runs the unified graph-construction, no-graph MLP and GNN evaluation protocols.

pip install tdg-bench

Important UMLS restriction

The UMLS-NCI graph and its generated variants are not included. They must be reconstructed locally by users holding a valid UMLS licence. See LICENSE_DATA.md and THIRD_PARTY_NOTICES.md for file-level terms.

Release

This directory is the data payload for TDGBench v2.0.0, extending the Biomedical benchmark from v1.0.0 with the DBpedia-174 domain. Consumers should pin the Hugging Face revision tag rather than an unversioned branch.