jknafou/trec-ragtime-2026

TREC RAGTIME 2026 — sentence and passage renderings

A sentence-level view of the TREC RAGTIME 2026 news collection, with two English machine translations of every non-English sentence and the passage boundaries used for retrieval. Derived from trec-ragtime/ragtime2.

Pipeline, experiment design, run configurations and reproduction steps: github.com/jknafou/trec-ragtime-2026

What is in here

Config	Splits	Rows	Contents
sentences	eng, spa, rus, zho	88,719,200	The segmentation, with character offsets into the source document
translations_nllb	spa, rus, zho	55,919,788	Every non-English sentence in English, by NLLB-200-3.3B
translations_opus	spa, rus, zho	55,919,788	The same sentences by the small Helsinki OPUS-MT models
passages	eng, spa, rus, zho	9,941,840	Retrieval units, each a contiguous run of sentences

Document and sentence ids are shared across every config, and unchanged from the parent, so a passage can be read as native text or as either translation without any alignment step.

Quick start

from datasets import load_dataset

repo = "jknafou/trec-ragtime-2026"
sentences = load_dataset(repo, "sentences", split="zho")
print(sentences[0])

Rebuilding a document's passage split, and the text of each passage:

from datasets import load_dataset

repo, split = "jknafou/trec-ragtime-2026", "zho"
passages = load_dataset(repo, "passages", split=split)
sentences = load_dataset(repo, "sentences", split=split)

doc = passages[0]["document_id"]
by_id = {r["sentence_id"]: r for r in sentences if r["document_id"] == doc}

for p in (r for r in passages if r["document_id"] == doc):
    members = [by_id[sid] for sid in p["sentence_ids"]]
    parts = [members[0]["text"]]
    for prev, cur in zip(members, members[1:]):
        if cur["char_start"] > prev["char_end"]:
            parts.append(" ")
        parts.append(cur["text"])
    print(p["passage_id"], len(members), "".join(parts))

The space goes in only where the offsets leave a gap: Chinese sentences often abut with no separator, so a plain join would corrupt them. Passages overlap, so a sentence can belong to more than one. For an English rendering, join the member translations with a single space.

Licence and attribution

Released under CC-BY-SA-4.0, inherited from the parent collection, which requires attribution, a statement of changes, and same-licence redistribution. The changes are sentence segmentation, passage grouping, and machine translation of the non-English sentences.

The translation models carry their own terms, which apply to the models and not to this text: NLLB-200-3.3B is CC-BY-NC-4.0, and the OPUS-MT checkpoints are Apache-2.0 and CC-BY-4.0.

Citation

A paper describing these runs is in preparation. This section will carry the reference once it is published.

% to be added