Awesome Search KG

ADE Corpus V2

3 min read

ADE Corpus V2

A biomedical benchmark of sentences drawn from medical case reports, labelled for whether they describe an adverse drug event. The corpus originates with Gurulingappa et al., Development of a benchmark corpus to support the automatic extraction of drug-related adverse effects from medical case reports, Journal of Biomedical Informatics 45(5): 885–892 (2012), and is redistributed on Hugging Face.

🔗 https://huggingface.co/datasets/ade-benchmark-corpus/ade_corpus_v2

Shape

The ade_corpus_v2_classification configuration is a sentence-level binary task: each row is one sentence, labelled ade_related or not_related. It carries 23,516 rows, reducing to 20,895 after normalization and deduplication. Positives are a minority — roughly one sentence in five in the samples drawn from it.

Two properties worth knowing before using it:

  • There are no article identifiers in the classification table. Deduplicating sentences cannot stop two sentences from the same case report landing on opposite sides of a train/test split, so a nominally held-out set is not held out at the document level.
  • The Hugging Face dataset card lists the license as unknown.

Pinning the revision matters; work in this vault uses 4ba01c7.

Why It Appears Here

Not a search dataset. It appears in this vault because Adapting Jev to Your Domain with GEPA uses it as the testbed for the first independent measurement of Jev’s calibration, and the task shape transfers directly to relevance work: a cheap classifier over a large candidate pool, an imbalanced positive class, a threshold you choose, and a human queue behind it.

That structure is why the paper’s review-policy tables — how much of the queue can be skipped, and how many positives get deferred doing it — are the reusable part, rather than the F1 numbers. It is Staged Judging on a literature-screening workload, and the same arithmetic governs a relevance-judging backlog.

Headline measurements taken on it: Jev at its default prompt reached 95.1% recall at 53.2% precision on a 300-sentence test set, against 67.2%/58.6% for a threshold-tuned TF-IDF baseline fit on 20,395 labelled sentences — while scoring substantially worse on every probability metric. A GEPA-optimized prompt then moved F1 from 69.1% to 79.7% on a fresh 300-sentence split.

Graph View

Backlinks