Hyukkyu/train-rlhn-msmarco

RLHN MS MARCO passage — Training, unified schema

The msmarco subset of rlhn/rlhn-680K (RLHN: the BGE retrieval training collection with its false hard negatives relabelled by a GPT-4o-mini → GPT-4o cascade; Thakur et al., 2025), reshaped into the strict schema shared by every dataset in this collection.

Source	rlhn/rlhn-680K @ d5323fad7aea, subset msmarco
Task	web query → passage
Domain · languages	web · eng
Queries / documents / qrels	455,177 / 5,390,108 / 1,344,758
Qrels per query	min 1 · mean 2.95 · max 15
Score values	2 ×455,119 · 1 ×889,639 (2: RLHN's original positive, 1: a hard negative RLHN relabelled as relevant)
Layout	queries · corpus · qrels · hard-negatives · teacher-scores, split train
Splits	corpus: train · hard-negatives: train · qrels: train · queries: train · teacher-scores: train
Hard negatives	sources: dataset · 10,515,228 rows
Teacher scores	jinaai/jina-reranker-v3.5 · 10,970,405 rows (positives included)
Ids	sha1(text)[:20]; identical texts collapse to one document / query
License	cc-by-sa-4.0

Schema

config	columns	rules
queries	id: string, text: string	ids unique and non-empty; every query has ≥ 1 qrel
corpus	id: string, title: string, text: string	title is always present ("" when the source has none)
qrels	query-id: string, corpus-id: string, score: int32	referential integrity to both tables; no duplicate pairs; no floats
hard-negatives	query-id: string, corpus-id: string, rank: int32, source: string	one row per negative; (query-id, corpus-id, source) unique; never a labelled positive of the same query
teacher-scores	query-id: string, corpus-id: string, teacher: string, score: float32	one row per scored pair (positives included); a row means scored — never a placeholder

Files are Parquet, sorted by id, zstd-compressed, sharded at 500 MB. Every rule above is checked before publishing; provenance.json records the source revision, what changed, and the output file hashes. teacher-scores holds no rows yet; datasets cannot return a 0-example split, so read that file with pyarrow / pandas until it is filled.

What changed from the source

• subset: the rows of rlhn/rlhn-680K whose subset is msmarco (MS MARCO passage in the BGE training collection)
• positives: each query's positive passages; RLHN's original positive at grade 2, the hard negatives its LLM cascade relabelled as relevant at grade 1
• negatives: RLHN's remaining hard negatives in RLHN's order (rank 1, 2, …), source = dataset
• decontaminated (exact queries): 4 queries whose normalised text equals an evaluation query (the 23 test sets and the 6 dev sets) were removed
• decontaminated (overlap audit): 271 passages that equal or nearly copy an evaluation document some evaluation query judges relevant, and 6 queries that nearly copy an evaluation query (word 13-grams for passages, 8-grams for queries; at least half shared with one evaluation text) were removed, and with them 14 queries in total; near copies of evaluation-corpus documents that no evaluation query judges relevant were kept
• text: leading and trailing whitespace stripped
• ids re-keyed to sha1(text)[:20]: 0 documents and 553 queries collapsed into identical texts
• not published: the order scores the local conversion uses in place of teacher scores (RLHN has no teacher)

Hard negatives and teacher scores

Filled by the owner's annotation pipeline (annotation=jina35-r4) for the train split of the query set(s) below; queries without a labelled positive are left out.

• Candidates: RLHN's own hard negatives only, scored as they are (no dense mining for this subset); rank is RLHN's order and source is dataset for every row.
• Teacher: jinaai/jina-reranker-v3.5, listwise: a query's positive and all of its candidates are scored together in one context of up to 32,768 tokens. score is the raw cosine score, one row per (query, positive) and per (query, candidate); a labelled negative that was also mined is scored once. Every document was cut to its first 1,024 reranker tokens before scoring (max_doc_tokens=1024). The positive scored is the query's training positive (its highest-graded qrel, ties by id); RLHN's other labelled positives are not scored. No filtering is applied to the tables.

configs	queries	hard negatives	teacher scores
hard-negatives · teacher-scores	455,177 (all)	10,515,228 (10,515,228 dataset)	10,970,405

from datasets import load_dataset
negatives = load_dataset("Hyukkyu/train-rlhn-msmarco", "hard-negatives", split="train")
scores    = load_dataset("Hyukkyu/train-rlhn-msmarco", "teacher-scores", split="train")

Load it

from datasets import load_dataset
queries   = load_dataset("Hyukkyu/train-rlhn-msmarco", "queries", split="train")
corpus    = load_dataset("Hyukkyu/train-rlhn-msmarco", "corpus", split="train")
qrels     = load_dataset("Hyukkyu/train-rlhn-msmarco", "qrels", split="train")
negatives = load_dataset("Hyukkyu/train-rlhn-msmarco", "hard-negatives", split="train")

License and attribution

RLHN is released under CC BY-SA 4.0, and so is this copy (share-alike). The underlying MS MARCO passage data carries its own terms. All credit belongs to the original authors; see rlhn/rlhn-680K. This repository is an independent repackaging. Please cite RLHN:

@misc{thakur2025relabel,
      title={Fixing Data That Hurts Performance: Cascading LLMs to Relabel Hard Negatives for Robust Information Retrieval},
      author={Nandan Thakur and Crystina Zhang and Xueguang Ma and Jimmy Lin},
      year={2025},
      eprint={2505.16967},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2505.16967},
}