ImpulseLeap/ru-retrieval-news

RU Retrieval News

A Russian-language news retrieval dataset: ~1.9M news documents and ~1.9M synthetically generated search queries, each paired with a precomputed dense embedding.

Built and maintained by Alexei Goncharov (ImpulseLeap).

Dataset summary

Two parquet files:

File	Rows	Description
corpus_final.parquet	1,925,391	News documents with text and embeddings
queries_final.parquet	1,900,815	Synthetic search queries with embeddings

Each document has exactly one corresponding query, matched by id. A small number of documents (~25k) have no query (failed generation) and are excluded from the queries file but kept in the corpus file, so the two files are not perfectly 1:1 in row count.

corpus_final.parquet columns

Column	Type	Description
id	string	Unique document id, prefixed by source (lenta-*, ria-*, qa-*)
text	string	Full raw article text (original casing/punctuation preserved)
symbols	int32	Character count of text
source	string	Original article URL (Lenta only) or source dataset tag (ria-news, qa_news_ru)
vector	fixed_size_list[4096]	Dense embedding of text

queries_final.parquet columns

Column	Type	Description
id	string	Matches the id of the source document in corpus_final.parquet
query	string	Synthetic natural-language search query for that document
vector	fixed_size_list[4096]	Dense embedding of query

Embeddings

All vectors were generated with Qwen/Qwen3-Embedding-8B via the DeepInfra API, encoding format float, then cast from float32 to float16 for storage. Empirically this loses negligible precision for retrieval purposes (cosine similarity between the float32 and float16 versions of the same vector is ~1.0 in spot checks).

Vectors are 4096-dimensional. No hard negatives are precomputed or shipped — the raw vectors are provided so that negative mining can be done with whatever embedding model, similarity threshold, or scale (5 negatives or 5,000) fits your training setup, via FAISS or similar over the vector column.

Query generation

Queries were generated with Gemini 2.5 Flash-Lite (Vertex AI Batch API), one query per document, prompted to produce a natural search-engine-style query a person might type to find that specific article (not a headline or summary).

Data sources

Documents were aggregated from three existing Russian-language news corpora:

Source	Prefix	Origin
Lenta	lenta-*	TopicNet/Lenta
ria-news	ria-*	ai-forever/ria-news-retrieval (via kaengreg/ria-news mirror), part of the ruMTEB benchmark suite
qa_news_ru	qa-*	AIR-Bench/qa_news_ru (mixed Russian-language news, not exclusively Russia-sourced)

License and content notice

This dataset (its structure, the synthetic queries, and the embeddings) is released under CC-BY-4.0. Please credit Alexei Goncharov / ImpulseLeap if you use it.

The underlying news article texts are aggregated from the public sources listed above for research purposes; they were not authored by the dataset creator, and their original copyright status varies by source and is not independently verified here. If you are a rights holder and want content removed, please open an issue on this repository or contact the author, and it will be handled promptly.

Usage

import pyarrow.parquet as pq

corpus = pq.read_table("corpus_final.parquet")
queries = pq.read_table("queries_final.parquet")

For large-scale use, prefer streaming by row group rather than loading the full table into memory — the corpus file alone contains ~30GB of uncompressed float16 vector data.

pf = pq.ParquetFile("corpus_final.parquet")
for batch in pf.iter_batches(batch_size=10_000):
    ...