datamaestro-ir 0.4.0

Datamaestro module for Information Retrieval datasets

Information Retrieval Datasets

This datamaestro plugin provides easy and systematic access to information retrieval datasets. It handles automated downloading and preparation of standard IR collections, exposes them through a typed Python API, and includes efficient document stores for fast text access (file, mmap, or in-memory).

Full documentation: datamaestro-ir.readthedocs.io

Available Datasets

Ad-hoc Retrieval

• TREC Ad-hoc (1–8), Robust 2004/2005 — classic TREC test collections over TIPSTER/AQUAINT corpora
• BEIR Benchmark — 15+ datasets: TrecCovid, NQ, ArguAna, Touché, ClimateFever, SciDocs, NFCorpus, HotpotQA, FiQA, Quora, DBpedia-Entity, FEVER, SciFact, CQADupStack (12 sub-forums)
• LoTTE — domain-specific retrieval across 6 domains (lifestyle, recreation, science, technology, writing, pooled) × dev/test × search/forum queries
• MS MARCO Passage & Document — passage ranking (8.8M passages) and document ranking (v1: 3.2M, v2: 12M documents)
• CORD-19 / TREC-COVID — COVID-19 research article retrieval (192K documents)

Conversational Search

• TREC CaST 2019–2022 — conversational passage retrieval with decontextualized queries, tree-structured conversations (2022), and segmented passages
• iKAT 2023–2025 — interactive knowledge-seeking over ClueWeb22

Query Rewriting

• CANARD — context-aware query rewriting (train/dev/test)
• QReCC — question rewriting in conversational context (14K conversations, 81K QA pairs)
• OrConvQA — open-retrieval conversational QA over 11M Wikipedia passages

Knowledge Distillation & Training Data

• MS MARCO Ensemble/BERT Teacher — 40M triples with teacher scores
• rank-distillm — BM25/ColBERTv2/RankZephyr annotated passages
• MS MARCO Hard Negatives — hard negatives mined from multiple retrieval models
• Neural Ranking KD — knowledge distillation teacher scores
• LightOn embeddings-pre-training — 73-config variant family ai.lighton.embeddings_pre_training[...] + DenseON-LateON mGTE recipe ...denseon_lateon.

Base Document Collections

• TIPSTER (AP, FT, WSJ, ZIFF, …), AQUAINT, TREC CAR (29.8M paragraphs), WAPO v2/v4, KILT (42M Wikipedia articles)