nickprock/it-wiki-retrieval-synthetic-hn

Italian Synthetic Retrieval Dataset with Dense Hard Negatives

Dataset Summary

This dataset is a high-quality, synthetic Information Retrieval (IR) dataset for the Italian language. It is designed to train and fine-tune state-of-the-art embedding models and bi-encoders (e.g., using MultipleNegativesRankingLoss).

The dataset consists of exactly 50,000 synthetic search queries generated from approximately 25,000 unique passages extracted from the Italian Wikipedia. Furthermore, it is augmented with Dense Hard Negatives mined iteratively using a strong cross-encoder/bi-encoder baseline to provide a highly challenging contrastive learning environment.

Dataset Structure

The dataset contains 50,000 rows. Each row represents a training instance with the following fields:

• query (or anchor): A synthetically generated search query in Italian (minimum 5 words). It alternates between a specific question and a broad semantic search intent.
• positive: The ground-truth paragraph from Italian Wikipedia containing the exact context for the query.
• hard_negatives: A list of text passages that are semantically similar to the query but do not answer it (mined via Dense Retrieval).

Dataset Creation Pipeline

The creation of this dataset follows a modern, highly optimized pipeline to prevent common contrastive learning issues (like vector space anisotropy).

1. Source Data: ~25,000 paragraphs were sampled from the Italian Wikipedia dump (wikimedia/wikipedia, 20231101.it).
2. Synthetic Query Generation: We used the Qwen-2.5-7B LLM locally to act as a data engineer. For each paragraph, the LLM was prompted strictly via structured JSON schema (Pydantic) to generate two types of queries:
   • One specific question (domanda_specifica).
   • One broad semantic search intent (ricerca_semantica).
   • Constraint: Single-word keyword queries were explicitly forbidden to force the generation of long-tail semantic queries.
3. Dense Hard Negative Mining: Instead of relying on traditional lexical search (BM25), which often produces weak negatives, we employed a Dense Self-Mining approach. We used nickprock/multi-sentence-BERTino (V1) to encode the entire corpus and perform a semantic search for each query. The top highly-ranked documents that were not the actual positive passage were selected as Hard Negatives.

Intended Use

This dataset is plug-and-play for training models using the sentence-transformers library.

Training Recommendation: When training with CachedMultipleNegativesRankingLoss or MultipleNegativesRankingLoss, we highly recommend unpacking the hard_negatives list into explicit columns (e.g., negative_1, negative_2) and using BatchSamplers.NO_DUPLICATES to avoid in-batch collisions and false negatives.

Author

Created by Nicola Procopio.