fabianschmidt-cohere/rcp-ndcg-vidore-v3

RCP-nDCG · ViDoRe v3 (MTEB-style, with continuous relevance gains)

Evaluate with stock mteb (≥ 2.10.5). No fork is needed. This repo ships rcp_ndcg_tasks.py. It defines one mteb task per subset (8 tasks, e.g. Vidore3ComputerScienceRCPReranking), 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-vidore-v3", "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()                        # all subsets; or rcp.get_tasks(["hr", "energy"])
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 (Vidore3ComputerScienceRCPReranking), the recommended setting.
• rcp.get_tasks(view="retrieval"): open-corpus retrieval (Vidore3ComputerScienceRCPRetrieval), bi-encoders only; see Recommended use below.
• rcp.get_tasks(text_only=True): the OCR'ed-markdown text view (…RCPRerankingOCR), for text-only models; combine with view="retrieval" for its open-corpus variant.

The same reranking tasks are in mteb itself once PR #5516 is merged. 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 8 public ViDoRe v3 domains × 6 query languages form 48 subsets. Each subset gets the judged 150-page candidate pool and calibrated relevance probabilities from an LLM judge. The full provenance of how the pools and gains were made is included.

RCP-nDCG (nDCG over rubric-calibrated preferences) replaces binary 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: Qwen3.5-397B-A17B (NVFP4)

Every gain value comes from a single LLM judge, Qwen3.5-397B-A17B (NVFP4 quantisation). The judge rated each pooled page against a 5-criterion ladder. A 2PL IRT model was fitted to those ratings 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

It maps into (0, 1) and uses the fitted item_params (provenance/item_params.qwen35.json). The raw calibrated theta ships next to gain, so any other link can be recomputed without re-judging. ViDoRe v3 was judged by one judge (Qwen3.5-397B).

Releases and judge provenance

release	Hub tag	revision (pinned by mteb)	judge	calibration	status
1	v1-nvidia-qwen3.5-397b-a17b-nvfp4	b22323345cb5639944600d490969bcce9fba7943	nvidia/Qwen3.5-397B-A17B-NVFP4	2PL IRT, provenance/item_params.qwen35.json	superseded
2	v2-nvidia-qwen3.5-397b-a17b-nvfp4	this revision	nvidia/Qwen3.5-397B-A17B-NVFP4	2PL IRT, provenance/item_params.qwen35.json	current

• The data (corpus, queries, qrels with gain / theta, pools and provenance) was last changed in this revision, the commit tagged v2-nvidia-qwen3.5-397b-a17b-nvfp4. The same commit holds this card, provenance/releases.json and the rcp_ndcg_tasks.py of release 2. mteb and the reference results pin this commit by its sha; the tag names the same commit, so both load identical data. The v1 tag v1-nvidia-qwen3.5-397b-a17b-nvfp4 remains loadable and keeps the v1 gains (data commit adb6a2a8).
• 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 Relation to the paper); 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 the commit sha that mteb pins: rcp.get_tasks(revision="v1-nvidia-qwen3.5-397b-a17b-nvfp4").
• The same information, machine-readable: provenance/releases.json.

Recommended use: reranking

Each subset ships top_ranked, the judged 150-page pool the paper's rerankers scored. Loaded through mteb, a model reranks exactly this pool. This is the recommended setting, because every candidate has a gain:

• ndcg_at_10: standard nDCG over the upstream graded qrels (0/1/2).
• 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.

Layout

Subset names are {domain}__{language}, e.g. energy__french.

• domains: computer_science, energy, finance_en, finance_fr, hr, industrial, pharmaceuticals, physics
• languages: english, french, german, italian, portuguese, spanish

config	columns	notes
{subset}-corpus	image, id, text	upstream page images + OCR markdown, byte-identical copy; one copy per domain, shared by its 6 languages
{subset}-queries	id, text	the queries in that language
{subset}-qrels	query-id, corpus-id, score, gain, theta	score: upstream graded qrels (1, 2), unchanged; gain: calibrated relevance probability (see below)
{subset}-top_ranked	query-id, corpus-ids	the judged 150-page pool
provenance-pool	query-id, corpus-id, pool_rank, rrf_score, gt_injected	split = domain
provenance-first-stage	retriever, query-id, corpus-id, rank, score	embed_v4, bm25, octen, rrf_fused
provenance-runs	model, query-id, corpus-id, score, rank_by_score, stored_position	14 rerankers
provenance-judge-criteria	query-id, call_index, corpus-id, C1…C5, finish_reason	14.5M per-call binary criterion verdicts, the 2PL fit's raw observations

Gains live in the qrels. {subset}-qrels has one row per (query, document) in the judged pool. Every human-labelled page is in the pool (ground-truth injection, below):

• score: the human relevance label, unchanged (0 for pooled pages without a 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). Two caveats:

• A score of 0 does not mean a human judged the document non-relevant. For the original human qrels, filter score > 0; for the judged pool, filter theta not null.
• Statistics that count "annotated" documents treat the judged pool as annotated. This affects mteb's descriptive statistic unique_relevant_docs, and trec_eval's bpref / judged@k.

Ids use the upstream form (query-test-<n>, corpus-test-<n>). They join exactly with mteb/Vidore3<Domain>OCRRetrieval at the revision pinned in the mteb task file, and with vidore/vidore_v3_<domain>_mteb_format.

How the pool was built (fully reproducible)

1. First stage, top-150 each from three retrievers over the page OCR text:
   • Cohere Embed v4
   • BM25
   • Octen-Embedding-8B
2. Reciprocal Rank Fusion: unweighted, k = 60, fused in the order embed_v4, bm25, octen, top 150. Fusing provenance-first-stage reproduces rrf_fused exactly: scores to <1e-12 and the same id order.
3. Ground-truth injection: any qrel-positive page missed by the fused top-150 replaces the lowest-ranked non-relevant page. It is marked gt_injected = true with rrf_score = 0. That is 11,843 of 2,177,100 pool pairs. Without it, the gains could not cover every relevant page.
4. Judging + calibration: the judge rated every pool page against the 5 criteria (C1–C5). A 2PL model was fitted, and θ was calibrated per query.

Results: 14 rerankers under both tie rules

nDCG@10 in %, each reranker scoring the judged 150-page pool (top_ranked). RCP-nDCG@10 uses the gain column with linear gain; its ideal ranking takes all gains of the query. qrel-nDCG@10 uses the integer score with linear gain; its ideal ranking takes all positive qrels. The two tie rules differ only in how pages with identical reranker scores are ordered:

• (a) breaks ties by document id (score descending, then corpus-id descending), the trec_eval / pytrec_eval convention.
• (b) credits each tied page its tie group's mean gain, the expected nDCG over all orderings of the tie. This is the paper's rule (arXiv v2: every suite, both metrics), and mteb's ndcg_float_at_k uses it.

Two views follow. The paper scores each question once, in its native language (french for energy, finance_fr and physics; english for the other five domains). The mteb tasks, like mteb's other ViDoRe v3 tasks, average one score per language subset (6 per domain); five of the six subsets pose translated questions over the same pages. Mean is the unweighted mean over the 8 domains. All numbers are replayed from provenance-runs at this revision.

Native-language questions (the paper's unit; 2,419 questions)

RCP-nDCG@10 (%), rule (a): document-id tie-break

Reranker	Comp. Sci.	Energy	Finance (en)	Finance (fr)	HR	Industrial	Pharma.	Physics	Mean
CTXL-RR-1B	85.2	88.1	82.5	82.2	85.3	81.5	83.7	86.8	84.4
CTXL-RR-2B	82.9	89.8	85.7	83.7	87.9	85.0	85.6	88.1	86.1
CTXL-RR-6B	79.0	89.0	86.5	85.5	88.0	85.1	84.7	88.1	85.7
Jina-RR-v3	73.3	70.5	67.8	64.2	69.7	74.3	76.0	77.0	71.6
Qwen3-RR-0.6B	82.4	79.6	74.6	66.6	73.1	76.3	77.8	83.4	76.7
Qwen3-RR-4B	87.7	86.6	79.5	80.4	82.7	81.1	83.9	88.2	83.8
Qwen3-RR-8B	88.4	89.3	82.6	85.1	85.3	84.1	85.9	89.4	86.3
RR4-Fast	89.9	90.1	87.2	84.7	88.4	85.1	86.7	88.3	87.6
RR4-Pro	90.9	93.0	88.5	90.3	90.4	87.9	89.1	91.8	90.2
Voyage2.5-lt	91.7	90.6	86.8	85.2	89.8	87.7	88.7	87.0	88.4
Voyage2.5	93.3	93.3	86.5	87.4	91.7	89.1	90.3	91.1	90.4
zerank-1-sm	93.3	91.0	88.2	86.7	90.1	87.8	88.1	90.9	89.5
zerank-1	94.2	93.3	88.8	89.6	91.6	89.1	90.9	93.4	91.4
zerank-2	94.3	93.6	87.6	89.6	91.9	89.4	91.2	93.8	91.4

RCP-nDCG@10 (%), rule (b): group-mean gain for ties

Reranker	Comp. Sci.	Energy	Finance (en)	Finance (fr)	HR	Industrial	Pharma.	Physics	Mean
CTXL-RR-1B	85.1	88.0	82.5	82.2	85.3	81.5	83.7	86.8	84.4
CTXL-RR-2B	82.9	89.9	85.7	83.7	87.9	85.0	85.6	88.1	86.1
CTXL-RR-6B	79.0	89.0	86.5	85.5	88.0	85.1	84.7	88.1	85.7
Jina-RR-v3	73.3	70.5	67.8	64.2	69.7	74.3	76.0	77.0	71.6
Qwen3-RR-0.6B	82.5	79.7	75.0	67.2	72.5	76.4	78.3	83.3	76.9
Qwen3-RR-4B	87.7	86.5	79.5	80.4	82.7	81.0	84.1	88.2	83.7
Qwen3-RR-8B	88.4	89.2	82.6	85.1	85.3	84.1	86.0	89.3	86.2
RR4-Fast	89.9	90.1	87.2	84.7	88.4	85.1	86.8	88.3	87.6
RR4-Pro	90.8	93.0	88.6	90.2	90.4	87.9	89.1	91.7	90.2
Voyage2.5-lt	91.6	90.7	86.8	85.2	89.8	87.7	88.6	87.0	88.4
Voyage2.5	93.3	93.3	86.5	87.5	91.8	89.1	90.3	91.1	90.4
zerank-1-sm	93.2	90.9	88.2	86.7	90.1	87.8	88.1	90.9	89.5
zerank-1	94.2	93.3	88.8	89.6	91.6	89.1	90.9	93.4	91.4
zerank-2	94.3	93.6	87.6	89.6	91.9	89.4	91.2	93.8	91.4

qrel-nDCG@10 (%), rule (a): document-id tie-break

Reranker	Comp. Sci.	Energy	Finance (en)	Finance (fr)	HR	Industrial	Pharma.	Physics	Mean
CTXL-RR-1B	77.7	68.9	69.0	49.7	65.5	57.4	66.4	49.1	63.0
CTXL-RR-2B	76.8	71.6	71.7	53.2	68.1	61.1	67.4	49.0	64.9
CTXL-RR-6B	74.2	71.9	73.8	55.2	69.1	60.7	68.1	51.2	65.5
Jina-RR-v3	74.1	65.8	63.6	41.5	57.2	57.3	67.9	45.3	59.1
Qwen3-RR-0.6B	74.1	53.3	57.1	33.1	47.1	48.5	59.5	47.8	52.5
Qwen3-RR-4B	80.3	68.4	67.7	50.9	61.8	56.7	66.9	51.3	63.0
Qwen3-RR-8B	80.8	71.0	70.2	55.7	65.9	59.6	68.7	52.7	65.6
RR4-Fast	82.0	70.8	74.5	53.6	68.5	59.7	69.4	48.8	65.9
RR4-Pro	83.2	73.7	75.5	57.9	70.8	61.3	71.3	51.5	68.1
Voyage2.5-lt	83.8	73.2	74.7	57.2	69.5	62.0	70.2	48.8	67.4
Voyage2.5	85.1	73.8	75.2	58.4	71.8	61.9	71.6	50.3	68.5
zerank-1-sm	84.3	69.6	72.9	52.8	66.7	58.6	67.4	48.6	65.1
zerank-1	84.0	71.6	75.1	55.5	68.7	60.4	69.7	49.0	66.7
zerank-2	84.1	72.4	75.1	55.8	68.8	60.7	69.7	49.0	67.0

qrel-nDCG@10 (%), rule (b): group-mean gain for ties

Reranker	Comp. Sci.	Energy	Finance (en)	Finance (fr)	HR	Industrial	Pharma.	Physics	Mean
CTXL-RR-1B	77.7	68.9	69.1	49.7	65.4	57.4	66.4	49.2	63.0
CTXL-RR-2B	76.7	71.6	71.7	53.2	68.1	61.1	67.5	49.0	64.9
CTXL-RR-6B	74.4	71.8	73.8	55.2	69.1	60.6	68.2	51.2	65.5
Jina-RR-v3	74.1	65.8	63.6	41.5	57.2	57.3	67.9	45.3	59.1
Qwen3-RR-0.6B	74.4	53.3	57.3	34.2	45.2	49.7	60.0	47.9	52.8
Qwen3-RR-4B	80.4	67.9	67.4	51.0	61.7	57.8	66.4	51.0	62.9
Qwen3-RR-8B	80.8	70.6	70.3	55.5	65.5	60.0	68.6	52.3	65.4
RR4-Fast	81.9	70.7	74.5	53.6	68.5	59.7	69.5	48.8	65.9
RR4-Pro	83.2	73.7	75.5	57.9	70.8	61.3	71.4	51.5	68.2
Voyage2.5-lt	83.8	73.2	74.6	57.3	69.6	62.1	70.2	49.0	67.5
Voyage2.5	85.1	74.0	75.1	58.5	71.9	62.1	71.6	50.3	68.6
zerank-1-sm	84.3	69.5	72.8	52.6	66.8	58.7	67.5	48.5	65.1
zerank-1	84.0	71.7	75.1	55.4	68.7	60.4	69.7	48.9	66.7
zerank-2	84.2	72.4	75.1	55.8	68.9	60.7	69.8	49.0	67.0

Mean over the 6 language subsets (the mteb task score)

RCP-nDCG@10 (%), rule (a): document-id tie-break

Reranker	Comp. Sci.	Energy	Finance (en)	Finance (fr)	HR	Industrial	Pharma.	Physics	Mean
CTXL-RR-1B	83.3	84.2	78.1	79.2	80.2	73.3	80.7	84.7	80.5
CTXL-RR-2B	85.5	87.2	83.3	81.5	85.8	81.5	85.3	86.8	84.6
CTXL-RR-6B	80.6	85.4	84.4	83.3	86.0	81.5	83.8	85.4	83.8
Jina-RR-v3	71.7	70.1	64.8	63.9	65.8	68.1	75.3	76.2	69.5
Qwen3-RR-0.6B	78.6	77.0	68.5	63.1	70.0	69.5	74.5	82.3	72.9
Qwen3-RR-4B	87.2	85.2	77.6	79.1	82.1	79.4	83.3	88.1	82.8
Qwen3-RR-8B	88.4	88.2	81.1	83.9	84.6	82.0	85.4	89.1	85.3
RR4-Fast	87.8	88.7	84.5	83.4	84.0	79.7	85.3	88.4	85.2
RR4-Pro	90.8	92.5	88.0	90.2	89.1	86.1	88.2	92.1	89.6
Voyage2.5-lt	88.3	89.3	83.1	83.6	86.7	81.6	87.4	86.3	85.8
Voyage2.5	92.0	92.8	84.9	87.2	90.2	85.7	89.7	90.8	89.2
zerank-1-sm	91.2	90.1	85.1	85.5	87.1	83.0	86.3	90.7	87.4
zerank-1	92.8	92.7	86.9	89.0	90.4	86.7	89.5	92.9	90.1
zerank-2	93.2	92.9	86.2	89.2	90.6	86.6	89.6	93.4	90.2

RCP-nDCG@10 (%), rule (b): group-mean gain for ties

Reranker	Comp. Sci.	Energy	Finance (en)	Finance (fr)	HR	Industrial	Pharma.	Physics	Mean
CTXL-RR-1B	83.3	84.2	78.1	79.1	80.3	73.3	80.7	84.7	80.5
CTXL-RR-2B	85.5	87.2	83.3	81.5	85.8	81.5	85.3	86.7	84.6
CTXL-RR-6B	80.6	85.4	84.4	83.3	86.0	81.5	83.8	85.4	83.8
Jina-RR-v3	71.7	70.1	64.8	63.9	65.8	68.1	75.3	76.2	69.5
Qwen3-RR-0.6B	78.7	77.3	69.0	64.3	69.5	69.6	75.2	82.1	73.2
Qwen3-RR-4B	87.2	85.0	77.7	79.0	82.1	79.4	83.5	88.0	82.7
Qwen3-RR-8B	88.4	88.1	81.2	84.0	84.6	82.1	85.4	89.0	85.3
RR4-Fast	87.8	88.7	84.6	83.4	84.0	79.7	85.3	88.4	85.2
RR4-Pro	90.8	92.5	88.0	90.2	89.1	86.1	88.2	92.1	89.6
Voyage2.5-lt	88.2	89.3	83.2	83.6	86.7	81.6	87.4	86.3	85.8
Voyage2.5	92.0	92.8	84.9	87.2	90.2	85.6	89.7	90.8	89.1
zerank-1-sm	91.2	90.1	85.1	85.5	87.1	83.0	86.3	90.7	87.4
zerank-1	92.8	92.7	86.9	89.0	90.4	86.7	89.5	92.9	90.1
zerank-2	93.2	92.9	86.2	89.2	90.6	86.6	89.6	93.4	90.2

qrel-nDCG@10 (%), rule (a): document-id tie-break

Reranker	Comp. Sci.	Energy	Finance (en)	Finance (fr)	HR	Industrial	Pharma.	Physics	Mean
CTXL-RR-1B	76.4	65.0	61.9	45.4	57.9	48.2	62.8	47.9	58.2
CTXL-RR-2B	76.9	69.3	67.3	50.6	64.2	56.8	66.1	47.4	62.3
CTXL-RR-6B	74.4	69.0	68.8	52.1	65.9	57.0	66.6	49.6	62.9
Jina-RR-v3	71.0	62.0	55.2	40.2	50.4	50.1	65.1	44.9	54.9
Qwen3-RR-0.6B	69.8	50.8	48.7	30.3	43.4	43.4	55.8	46.7	48.6
Qwen3-RR-4B	78.5	67.2	63.1	49.9	61.3	54.5	65.1	50.8	61.3
Qwen3-RR-8B	80.4	69.4	66.7	53.8	63.5	57.2	67.5	52.2	63.8
RR4-Fast	79.3	68.7	68.6	52.3	62.3	54.1	67.2	48.4	62.6
RR4-Pro	82.0	73.2	73.3	57.5	68.1	60.1	70.2	51.3	67.0
Voyage2.5-lt	80.3	71.4	68.0	54.1	65.1	55.2	68.1	47.7	63.7
Voyage2.5	83.9	73.3	71.6	57.5	68.8	58.0	69.6	49.5	66.5
zerank-1-sm	81.9	68.6	67.3	51.0	63.1	53.4	64.8	47.7	62.2
zerank-1	82.9	71.2	70.7	54.5	66.7	57.4	67.8	47.8	64.9
zerank-2	83.0	71.6	70.9	55.2	67.0	57.2	67.7	48.3	65.1

qrel-nDCG@10 (%), rule (b): group-mean gain for ties

Reranker	Comp. Sci.	Energy	Finance (en)	Finance (fr)	HR	Industrial	Pharma.	Physics	Mean
CTXL-RR-1B	76.4	65.0	61.9	45.4	57.9	48.3	62.8	47.9	58.2
CTXL-RR-2B	76.9	69.3	67.3	50.6	64.3	56.8	66.2	47.4	62.3
CTXL-RR-6B	74.4	69.0	68.8	52.1	66.0	57.0	66.6	49.6	62.9
Jina-RR-v3	71.0	62.0	55.2	40.2	50.4	50.1	65.1	44.9	54.9
Qwen3-RR-0.6B	69.9	51.2	49.4	31.9	42.4	44.0	56.3	46.6	49.0
Qwen3-RR-4B	78.6	66.4	62.9	49.8	61.2	55.2	65.0	50.5	61.2
Qwen3-RR-8B	80.3	69.0	66.7	53.6	63.4	57.5	67.3	51.8	63.7
RR4-Fast	79.3	68.7	68.6	52.3	62.3	54.0	67.2	48.5	62.6
RR4-Pro	82.0	73.2	73.3	57.5	68.1	60.1	70.3	51.3	67.0
Voyage2.5-lt	80.3	71.4	68.0	54.1	65.1	55.2	68.1	47.8	63.7
Voyage2.5	83.9	73.4	71.6	57.4	68.9	58.1	69.6	49.5	66.5
zerank-1-sm	81.9	68.5	67.3	51.0	63.2	53.4	64.9	47.7	62.2
zerank-1	82.9	71.2	70.8	54.5	66.7	57.4	67.9	47.9	64.9
zerank-2	83.0	71.7	70.9	55.2	67.0	57.2	67.7	48.4	65.1

Relation to the paper. These are the arXiv v2 numbers: the paper's ViDoRe table uses rule (b) on the native-language questions (2,419), which is the native-language view under (b). Release 2 corrects how the judge's tournament (Stage A) 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 6,051 of 3,122,771 tournament windows, of which v2 recovered 6,024. A tournament window now enters the fit only when the answer gives a complete ranking and a score for every document: release 2 leaves out the 27 answers it still cannot read and 2,322 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 (11 of 1,451,399 rubric windows), which is why provenance-judge-criteria has 99 fewer rows than in release 1. The 2PL fit was then redone, so the RCP-nDCG@10 cells change; qrel-nDCG@10 does not depend on the judge and is unchanged. Rule (a) gives the same view within 0.7 points on RCP-nDCG@10 (32 of 126 printed values differ) and within 1.9 points on qrel-nDCG@10 (64 of 126).

Native language vs the mteb task score. The mteb score averages the six language subsets (14,514 questions). Under rule (b), it lies on average 1.9 RCP-nDCG@10 points below the native-language number (per cell −8.2 to +2.6), and 2.7 qrel-nDCG@10 points below it (−9.1 to +0.1).

Effect of the tie rule. Per cell, the document-id rule (a) differs from the paper's rule (b) by at most 0.64 points on RCP-nDCG@10 in the native view, and by at most 1.2 points in the six-language view. On average, (a) minus (b) is −0.01 and −0.02 points. On qrel-nDCG@10 the maxima are 1.81 and 1.62 points.

Rerankers (hub model in provenance-runs): CTXL-RR-1B = ctxl-rerank-v2-instruct-multilingual-1b, CTXL-RR-2B = ctxl-rerank-v2-instruct-multilingual-2b, CTXL-RR-6B = ctxl-rerank-v2-instruct-multilingual-6b, Jina-RR-v3 = jina-reranker-v3, Qwen3-RR-0.6B = Qwen3-Reranker-0.6B, Qwen3-RR-4B = Qwen3-Reranker-4B, Qwen3-RR-8B = Qwen3-Reranker-8B, RR4-Fast = rerank-v4.0-fast, RR4-Pro = rerank-v4.0-pro, Voyage2.5-lt = rerank-2.5-lite, Voyage2.5 = rerank-2.5, zerank-1-sm = zerank-1-small, zerank-1 = zerank-1, zerank-2 = zerank-2.

Known data notes

• Language cannot be derived from row position. 5 of 8 domains have two batches of 6 language blocks. Always use the language of the upstream query, as the subsets here do.
• Some pages have byte-identical OCR text: 56 groups covering 181 pages, most in pharmaceuticals (26 groups, 84 pages). In 12 groups (pharmaceuticals 11, physics 1), identical pages carry different qrels.
• 264 pages have empty OCR text (industrial 205, finance_en 19, energy 15, hr 12, finance_fr 11, computer_science 2), so a text first stage can never retrieve them. None is in any pool or relevant to any query.

Licence and attribution

Derived from ViDoRe v3 (Loison et al., 2026), CC-BY-4.0. The corpus, queries and qrels are redistributed unchanged. The provenance-runs scores come from third-party reranker models and APIs, and the gains come from an LLM judge's outputs; check those providers' terms before any public or commercial reuse. Please cite ViDoRe v3 alongside RCP-nDCG:

@article{loison2026vidorev3comprehensiveevaluation,
  author = {António Loison and Quentin Macé and Antoine Edy and Victor Xing and Tom Balough and Gabriel Moreira and Bo Liu and Manuel Faysse and Céline Hudelot and Gautier Viaud},
  title = {ViDoRe V3: A Comprehensive Evaluation of Retrieval Augmented Generation in Complex Real-World Scenarios},
  journal = {arXiv preprint arXiv:2601.08620},
  year = {2026},
}

Please also cite 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.