README.md

metadata

license: cc-by-4.0
task_categories:
  - sentence-similarity
  - text-retrieval
language:
  - en
tags:
  - veterinary
  - medical
  - sentence-transformers
  - hard-negative-mining
size_categories:
  - n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: anchor
      dtype: string
    - name: positive
      dtype: string
    - name: negative_1
      dtype: string
    - name: negative_2
      dtype: string
    - name: negative_3
      dtype: string
    - name: negative_4
      dtype: string
    - name: negative_5
      dtype: string
    - name: id
      dtype: string
    - name: journal
      dtype: string
    - name: scores
      list: float64
  splits:
    - name: train
      num_bytes: 25871136
      num_examples: 4212
  download_size: 13300907
  dataset_size: 25871136

PMC Veterinary N-Tuples Dataset

Dataset Description

4,210 n‑tuples of (anchor, positive, 5 hard negatives) extracted from open‑access PubMed Central veterinary papers. Each anchor (title + background/objective) is paired with its own conclusions (positive) and five carefully mined hard negatives from different journals, ensuring no data leakage. The dataset is ready for training bi‑encoder models with MultipleNegativesRankingLoss.

Source Code & Reproducibility

• 💻 GitHub: petVet-dataset-creation-nhci
• 📓 Kaggle: petvet-dataset-creation

Dataset Structure

Each row has:

• anchor: string – research context (title + background/objective)
• positive: string – results/conclusions of the same paper
• negative_1 to negative_5: strings – hard negatives from other papers, pre‑filtered with PMID‑ and journal‑level masking
• id: string – PubMed ID (PMID) of the anchor paper
• journal: string – journal name

Mining Statistics (brief)

Metric	Value
Positive mean cosine similarity	0.70
Negative mean cosine similarity	0.49
Pos–Neg similarity gap	0.21
Cross‑journal purity	100% (no same‑journal negatives)

Intended Use

Fine‑tuning Sentence‑Transformers for veterinary/biomedical information retrieval. The dataset natively supports MultipleNegativesRankingLoss and can be used directly with the SentenceTransformerTrainer.

Quick Start

from datasets import load_dataset
dataset = load_dataset("dorrito-dev/pmc-vet-ntuples")