Datasets:

dorrito-dev
/

dog-breed-ntuple-retrieval

language:
  - en
license: cc-by-sa-4.0
task_categories:
  - text-retrieval
  - sentence-similarity
task_ids:
  - semantic-similarity-classification
tags:
  - dogs
  - dog-breeds
  - retrieval
  - n-tuple
  - hard-negatives
  - sentence-transformers
  - wikipedia
pretty_name: Dog Breed N-Tuple Retrieval Dataset
size_categories:
  - 1K<n<10K
source_datasets:
  - Wikipedia (en)

Dog Breed N-Tuple Retrieval Dataset

A hard-negative-mined n-tuple dataset for training and evaluating sentence embedding / bi-encoder retrieval models on dog-breed knowledge.

Dataset construction

Source: English Wikipedia articles for 60+ dog breeds (popular/AKC breeds + recognized Indian breeds).
Sections captured: Appearance, Temperament, Health, History, Care, Characteristics, and article summaries — extracted recursively so nested subsections are not missed.
Anchor generation: Up to 2 paraphrased query templates per passage (e.g. "What is the temperament of the Labrador Retriever?").
Hard-negative mining: FAISS-backed TopK mining with sentence-transformers/all-MiniLM-L6-v2, NV-Retriever-style positive-aware margin (relative_margin=0.05), and a cross-breed metadata mask that forbids negatives from the same breed as the anchor.
Train / eval split: held out by breed (not by row) so the eval split measures generalization to unseen breed names.

Column schema

Column	Description
`anchor`	Natural-language query about a breed
`positive`	Wikipedia passage that answers the query
`negative_1` … `negative_5`	Hard-negative passages from different breeds

Splits

Split	Rows	Breeds
train	see dataset	~85 % of breeds
eval	see dataset	~15 % of breeds (held out by breed)

Usage

from datasets import load_dataset
from sentence_transformers import SentenceTransformer
from sentence_transformers.losses import MultipleNegativesRankingLoss

ds = load_dataset("dorrito-dev/dog-breed-ntuple-retrieval")
train = ds["train"]

License

Derived from Wikipedia content, which is licensed under CC BY-SA 4.0.