Kushtrim/bujku_topics_pairs

Bujku Topics Pairs

Dataset of (topic → text) pairs derived from the Bujku archive, intended for embedding finetuning / retrieval training.

Dataset summary

Each example is a pair:

• anchor: a topic string (short text)
• positive: the associated full_text (longer text)

For each source JSON page/article, every topic in topics[] becomes one row where positive is the page/article text.

Data fields

The dataset contains the following columns:

• anchor (string): topic string.
• positive (string): article/page text associated with the topic.
• positive_token_length (int): token length of positive.

Intended use

• Semantic search / retrieval training (e.g., MultipleNegativesRankingLoss with in-batch negatives)
• Query ↔ passage similarity evaluation
• Clustering / deduplication experiments (after embedding)

Token length notes

positive_token_length is useful to:

• choose a training max_seq_length
• filter out examples that would be heavily truncated

Some training scripts may drop this column after filtering to keep only anchor/positive during training.

Loading the dataset

Hosted as:

• Kushtrim/bujku_topics_pairs

import os
from datasets import load_dataset

ds = load_dataset("Kushtrim/bujku_topics_pairs", split="train", token=os.environ.get("HF_TOKEN"))
print(ds.column_names, ds.num_rows)

Limitations and considerations

• Source domain is news-style text; performance may vary on other domains.
• OCR artifacts and noise may be present in positive.
• Duplicate or highly repetitive topics can exist across pages/issues.