Dataset Viewer
Auto-converted to Parquet Duplicate
query-id
string
call_index
int64
corpus-id
string
C1
int64
C2
int64
C3
int64
C4
int64
C5
int64
finish_reason
string
855410
0
2945339
0
0
0
0
0
stop
855410
0
2164297
0
0
0
0
0
stop
855410
0
7560274
0
0
0
0
0
stop
855410
0
5008542
0
0
0
0
0
stop
855410
0
4151850
0
0
0
0
0
stop
855410
0
7128710
0
0
0
0
0
stop
855410
0
367832
0
0
0
0
0
stop
855410
0
4810235
0
0
0
0
0
stop
855410
0
6425421
0
0
0
0
0
stop
855410
0
1732923
0
0
0
0
0
stop
855410
1
1041939
0
0
0
0
0
stop
855410
1
7238608
0
0
0
0
0
stop
855410
1
6704335
0
0
0
0
0
stop
855410
1
5465332
0
0
0
0
0
stop
855410
1
889438
0
0
0
0
0
stop
855410
1
8373834
0
0
0
0
0
stop
855410
1
7291107
0
0
0
0
0
stop
855410
1
476433
0
0
0
0
0
stop
855410
1
7972234
0
0
0
0
0
stop
855410
1
4263016
0
0
0
0
0
stop
855410
2
983642
0
0
0
0
0
stop
855410
2
681986
0
0
0
0
0
stop
855410
2
8671033
0
0
0
0
0
stop
855410
2
7238783
0
0
0
0
0
stop
855410
2
6707516
0
0
0
0
0
stop
855410
2
7140273
0
0
0
0
0
stop
855410
2
5210914
0
0
0
0
0
stop
855410
2
4299367
0
0
0
0
0
stop
855410
2
7509177
0
0
0
0
0
stop
855410
2
515982
0
0
0
0
0
stop
855410
3
2850382
0
0
0
0
0
stop
855410
3
8651775
1
1
1
1
0
stop
855410
3
4972697
0
0
0
0
0
stop
855410
3
5008545
0
0
0
0
0
stop
855410
3
4833539
0
0
0
0
0
stop
855410
3
3084026
0
0
0
0
0
stop
855410
3
5329993
0
0
0
0
0
stop
855410
3
6306434
0
0
0
0
0
stop
855410
3
5965635
0
0
0
0
0
stop
855410
3
8310088
0
0
0
0
0
stop
855410
4
6734486
0
0
0
0
0
stop
855410
4
6950985
0
0
0
0
0
stop
855410
4
1824280
0
0
0
0
0
stop
855410
4
8671031
0
0
0
0
0
stop
855410
4
8157624
0
0
0
0
0
stop
855410
4
5321528
0
0
0
0
0
stop
855410
4
6481240
0
0
0
0
0
stop
855410
4
3594261
0
0
0
0
0
stop
855410
4
7972231
0
0
0
0
0
stop
855410
4
8352927
0
0
0
0
0
stop
855410
5
4246164
0
0
0
0
0
stop
855410
5
7691815
0
0
0
0
0
stop
855410
5
4420015
0
0
0
0
0
stop
855410
5
5210919
0
0
0
0
0
stop
855410
5
181264
0
0
0
0
0
stop
855410
5
8468783
0
0
0
0
0
stop
855410
5
4103170
0
0
0
0
0
stop
855410
5
4146680
0
0
0
0
0
stop
855410
5
8361990
0
0
0
0
0
stop
855410
5
510943
0
0
0
0
0
stop
855410
6
1494095
0
0
0
0
0
stop
855410
6
7004260
0
0
0
0
0
stop
855410
6
4636144
0
0
0
0
0
stop
855410
6
681990
0
0
0
0
0
stop
855410
6
6947490
0
0
0
0
0
stop
855410
6
4708049
0
0
0
0
0
stop
855410
6
6165926
0
0
0
0
0
stop
855410
6
351153
0
0
0
0
0
stop
855410
6
2531289
0
0
0
0
0
stop
855410
6
3696341
0
0
0
0
0
stop
855410
7
6467175
0
0
0
0
0
stop
855410
7
1956109
0
0
0
0
0
stop
855410
7
7611461
0
0
0
0
0
stop
855410
7
8621844
0
0
0
0
0
stop
855410
7
681984
0
0
0
0
0
stop
855410
7
4151857
0
0
0
0
0
stop
855410
7
8651776
1
0
1
0
0
stop
855410
7
7210883
0
0
0
0
0
stop
855410
7
7031507
0
0
0
0
0
stop
855410
7
984533
0
0
0
0
0
stop
855410
8
197471
0
0
0
0
0
stop
855410
8
5723210
0
0
0
0
0
stop
855410
8
3249944
0
0
0
0
0
stop
855410
8
2147202
0
0
0
0
0
stop
855410
8
3557381
0
0
0
0
0
stop
855410
8
2519616
0
0
0
0
0
stop
855410
8
1075228
0
0
0
0
0
stop
855410
8
8087245
0
0
0
0
0
stop
855410
8
4406031
0
0
0
0
0
stop
855410
8
5381946
0
0
0
0
0
stop
855410
9
5210917
0
0
0
0
0
stop
855410
9
2785222
0
0
0
0
0
stop
855410
9
4635595
0
0
0
0
0
stop
855410
9
681985
0
0
0
0
0
stop
855410
9
1947769
1
1
1
1
0
stop
855410
9
1094091
0
0
0
0
0
stop
855410
9
2519618
0
0
0
0
0
stop
855410
9
8776516
0
0
0
0
0
stop
855410
9
1886069
0
0
0
0
0
stop
855410
9
8407524
0
0
0
0
0
stop
End of preview. Expand in Data Studio

RCP-nDCG · TREC-DL (MTEB-style, with continuous relevance gains)

Evaluate with stock mteb (≥ 2.0.1). No fork is needed. This repo ships rcp_ndcg_tasks.py. It defines one mteb task per subset (2 tasks, e.g. TRECDL2019RCPReranking), which reports the continuous-gain metric ndcg_float_at_10 alongside mteb's usual metrics:

import importlib.util
import mteb
from huggingface_hub import hf_hub_download

path = hf_hub_download("fabianschmidt-cohere/rcp-ndcg-trecdl", "rcp_ndcg_tasks.py", repo_type="dataset")
spec = importlib.util.spec_from_file_location("rcp_ndcg_tasks", path)
rcp = importlib.util.module_from_spec(spec)
spec.loader.exec_module(rcp)

tasks = rcp.get_tasks()                        # both subsets; or rcp.get_tasks(["trec_dl_2019"])
model = mteb.get_model("sentence-transformers/all-MiniLM-L6-v2")   # any mteb bi- or cross-encoder
results = mteb.evaluate(model, tasks=tasks)

Each view is its own task, so the task name in a results file says which one produced the score:

  • rcp.get_tasks(): reranking over the judged pool (TRECDL2019RCPReranking), the recommended setting.
  • rcp.get_tasks(view="retrieval"): open-corpus retrieval (TRECDL2019RCPRetrieval), bi-encoders only; see Recommended use below.

The reranking tasks are planned for mteb itself in a follow-up to PR #5516. When your mteb ships a task of the same name, it pins this release's commit by its sha. get_tasks(revision="<that sha>") returns the built-in task, and get_tasks() loads the same commit by its tag, so the numbers are identical either way.

Renamed (October 2026). Earlier versions of rcp_ndcg_tasks.py named the reranking tasks …RCPRetrieval. They are now …RCPReranking, like mteb's other reranking tasks, and …RCPRetrieval names the open-corpus view. Results computed with the earlier file under …RCPRetrieval names are reranking results; rename them to …RCPReranking. The earlier file also indexed only the pooled documents. That leaves dense-model scores unchanged but not those of models that use corpus statistics (e.g. BM25), so re-run those.

The two TREC Deep Learning Track passage ranking tasks (2019: 43 queries, 2020: 54 queries) over the MS MARCO passage collection, with the official NIST multi-graded relevance judgments (grades 0–3) and full provenance. Subsets are named trec_dl_2019 / trec_dl_2020, and the split is test, as in mteb's TRECDL2019 / TRECDL2020 tasks.

RCP-nDCG (nDCG over rubric-calibrated preferences) replaces integer relevance labels with calibrated relevance probabilities. An LLM judge compares pooled documents in a listwise tournament and checks each one against a rubric of five binary relevance criteria. A 2PL item-response model then calibrates those verdicts into a latent relevance theta per (query, document), and the relevance probability gain(theta) is the gain in nDCG.

⚠️ Gains come from ONE judge: Qwen3.6-27B-FP8

Every gain value comes from a single LLM judge, Qwen3.6-27B-FP8. Gains come from Qwen3.6-27B-FP8, a different judge from the NanoBEIR/BRIGHT/ViDoRe RCP-nDCG uploads (Qwen3.5-397B-A17B), so RCP scores here are not directly comparable to those suites. The judge answered five binary relevance criteria for each pooled document; a 2PL IRT model was fitted to those answers and calibrated. The gain is the paper's weighted_discrimination link:

gain(θ) = Σ_k γ_k · σ(γ_k (θ − β_k)) / Σ_k γ_k        ∈ (0, 1)
  • item_params for the gain are in provenance/item_params.qwen36-27b.json.
  • Every gain ships with its raw calibrated theta, so a different link can be applied without re-judging.

Releases and judge provenance

release Hub tag revision (pinned by mteb) judge calibration status
1 v1-qwen-qwen3.6-27b-fp8 9b78fbd9e86610a88f8d6c761974141b5d53ae8f Qwen/Qwen3.6-27B-FP8 2PL IRT, provenance/item_params.qwen36-27b.json superseded
2 v2-qwen-qwen3.6-27b-fp8 this revision Qwen/Qwen3.6-27B-FP8 2PL IRT, provenance/item_params.qwen36-27b.json current
  • The data (corpus, queries, qrels with gain / theta, pools and provenance) was last changed in this revision, the commit tagged v2-qwen-qwen3.6-27b-fp8. The same commit holds this card, provenance/releases.json and the rcp_ndcg_tasks.py of release 2. A pinned evaluation uses this commit's sha or the tag, which names the same commit, so both load identical data. The v1 tag v1-qwen-qwen3.6-27b-fp8 remains loadable and keeps the v1 gains (data commit 9b78fbd9).
  • A new judge or a recalibration is a new release: a new commit with the new gains, a new tag v<N>-<judge>, and new mteb task versions (<task>.v<N>) whose predecessors are marked superseded_by. Release 2 is a recalibration of the same judge's answers (see The paper's numbers under Data notes); it keeps the task names, since the mteb tasks are not merged yet. History is never rewritten or squashed, so every pinned revision stays loadable. rcp.get_tasks() loads release 2 by its tag. To evaluate a specific release, pass its tag or its commit sha: rcp.get_tasks(revision="v1-qwen-qwen3.6-27b-fp8").
  • The same information, machine-readable: provenance/releases.json.

Recommended use: reranking

{subset}-top_ranked is the judged candidate list, and every candidate has a gain. Loaded through mteb, a model reranks this list. This is the recommended setting:

  • ndcg_at_10: standard nDCG over the upstream NIST qrels (unchanged).
  • ndcg_float_at_10 (the main score): nDCG over the calibrated relevance probabilities, with linear gain. Documents tied on score are credited their tie group's mean gain.

Open-corpus retrieval works from the same data (view="retrieval", tasks …RCPRetrieval): the full upstream corpus is searched. The same metrics are reported, and the main score is still ndcg_float_at_10, but it becomes a pooled-judgment score: gains exist only for the judged pool, so a retrieved document outside it counts as gain 0. Compare open-corpus scores only with each other, not with the reranking scores.

Evaluation protocol

  • Judged pools. The candidates are the TREC-DL judged pools: the fused first-stage top-150 plus every NIST-judged document (pool sizes 198–585 per query). Every NIST-judged document is in the pool, so there are no unjudged human positives: every pooled document has a gain and a theta, and every human-labelled document has its score.
  • Grades. score is the official NIST relevance grade (0–3), unchanged from the upstream TREC-DL datasets. A score of 0 does not necessarily mean a human judged the document non-relevant: pooled documents outside the NIST judgments also carry score = 0 (filter theta not null to separate judge-scored documents, or score > 0 for the original human qrels).
  • Corpus text. The corpus is the MS MARCO passage collection as in ir_datasets msmarco-passage; its ids are identical to the upstream mteb TREC-DL datasets. That dataset's text carries a double-encoding (mojibake) error in 2,025,452 of 8,841,823 passages (22.9%), which is corrected here: the text in this repo is the text the judge saw.

Layout

config columns notes
{subset}-corpus id, title, text full MS MARCO passage corpus (8,841,823 passages), shared by both subsets
{subset}-queries id, text
{subset}-qrels query-id, corpus-id, score, gain, theta score: NIST grade, unchanged; gain: calibrated relevance probability (see above)
{subset}-top_ranked query-id, corpus-ids judged pool, in pool order
provenance-pool query-id, corpus-id, pool_rank, rrf_score, gt_injected, excluded, judged split = subset; excluded is always false here (no exclusions)
provenance-runs model, is_judge, query-id, corpus-id, score, rank_by_score, stored_position 14 rerankers + 2 judge strategies
provenance-judge-criteria query-id, call_index, corpus-id, C1…C5, finish_reason raw per-call criterion verdicts

Gains live in the qrels. {subset}-qrels has one row per (query, document) in the judged pool:

  • score: the NIST relevance grade (0–3), unchanged; 0 for pooled documents without a NIST label.
  • gain: the calibrated relevance probability, used by ndcg_float_at_k.
  • theta: the raw calibrated latent relevance behind gain.

Stock mteb reads only query-id, corpus-id, score. The extra score-0 rows change none of its retrieval scores (verified on every published run). A caveat: statistics that count "annotated" documents treat the judged pool as annotated, which affects mteb's descriptive statistic unique_relevant_docs and trec_eval's bpref / judged@k.

Query instructions equal the mteb task's prompt["query"] for both originals and are applied by mteb, so no -instruction config is shipped.

How the pool was built

  1. First stage: Cohere Embed v4, BM25 and Octen-Embedding-8B are fused with unweighted RRF (k = 60), keeping the top 150 (provenance-pool.rrf_score).
  2. Judged union: every NIST-judged document was added to the fused top-150 (the TREC-DL judged pools, up to 585 deep on 2019). Documents outside the fused top-150 have rrf_score = 0; qrel positives among them are marked gt_injected (1,844 on 2019, 1,566 on 2020).
  3. Judging: a listwise tournament, then pointwise binary criteria, then a 2PL fit.
    • Stage B rates every pooled document against five binary criteria (C1–C5), and a 2PL model is fitted to these verdicts.
    • The raw verdicts are in provenance-judge-criteria.

Data notes

  • The paper's numbers. mteb's ndcg_float_at_k gives tied documents their group's mean gain, so the number depends only on the reranker's scores. The paper's TREC-DL results on these gains (arXiv v2) use the same rule, and replaying provenance-runs reproduces them. arXiv v1 broke ties by the judged pool's order instead (the fused top 150 first, then the added NIST-judged passages). On these gains, that order moves a cell by up to 5.10 points on RCP-nDCG@10 and 8.63 points on nDCG@10 for Qwen3-RR-0.6B, whose scores have many ties, and by at most 0.62 and 1.95 points for every other reranker. Release 2 corrects how the judge's tournament answers were read: v1 replaced answers its parser could not read with rankings scraped from the response text and weighted them 5x in the fit, while v2 re-reads the judge's own answers. This concerned 8 of the 87,723 tournament windows of the shared fit (MS MARCO dev, TREC-DL 2019 and 2020), of which v2 recovered 6. A tournament window now enters the fit only when the answer gives a complete ranking and a score for every document: in the shared fit, release 2 leaves out the 2 answers it still cannot read and 217 answers that rank or score only some of their documents, which release 1 had fitted. Release 2 also applies the release rule to the rubric (Stage B) answers: an answer that does not give all five criteria for every document it shows is left out (28 of 10,961 rubric windows of the two subsets here), which is why provenance-judge-criteria has 251 fewer rows than in release 1.
  • is_judge = true runs in provenance-runs are the judge itself, not independent systems.
  • The 2PL calibration covers all 597 queries judged in this run (MS MARCO dev, TREC-DL 2019 and 2020) with one shared item-parameter fit; provenance/item_params.qwen36-27b.json holds the parameters used for the gains here.

Licence and attribution

Derived from the TREC-DL tasks in mteb (TRECDL2019, TRECDL2020), which are declared msr-la-nc (Microsoft Research License — non-commercial use only) in their mteb task metadata. Corpus, queries and qrels are redistributed unchanged at the pinned upstream revisions (the corpus text carries the encoding correction described above). The provenance-runs scores come from third-party reranker models and APIs, and the gains from an LLM judge's outputs, so those providers' terms also apply. Check the upstream terms before commercial use. Please cite MTEB and MS MARCO / TREC-DL alongside RCP-nDCG (arXiv:2609.35739):

@misc{schmidt2026rcp,
  title         = {Rubric-Calibrated Preferences: Cross-Query Calibration of LLM Judgments via Item Response Theory},
  author        = {Schmidt, Fabian David and Crisostomi, Donato and Lassance, Carlos and Reimers, Nils},
  year          = {2026},
  eprint        = {2609.35739},
  archiveprefix = {arXiv},
  primaryclass  = {cs.IR},
  url           = {https://arxiv.org/abs/2609.35739},
}

The blind human study and the external LLM-judge comparisons that validate RCP-nDCG are released in fabianschmidt-cohere/rcp-ndcg-external-validation.

Downloads last month
265

Paper for fabianschmidt-cohere/rcp-ndcg-trecdl