fabianschmidt-cohere/rcp-ndcg-external-validation

RCP-nDCG: external validation

This dataset holds the external validation of the paper "Rubric-Calibrated Preferences: Cross-Query Calibration of LLM Judgments via Item Response Theory" (Schmidt, Crisostomi, Lassance, Reimers, 2026; arXiv:2609.35739): the blind human study and the blind comparisons by external LLM judges. It contains:

• every grade, tie-break, review and verdict of the human study;
• every prompt and response of the three external LLM judges, reasoning traces included where they were stored.

The benchmark data are released separately:

• Paper: https://arxiv.org/abs/2609.35739
• Code: https://github.com/cohere-ai/rcp-ndcg
• Benchmark data: the RCP labels, calibration parameters, Stage A and Stage B judgments, and reranker runs of NanoBEIR, BRIGHT, ViDoRe v3 and TREC-DL are in the per-benchmark datasets, for example rcp-ndcg-nanobeir, rcp-ndcg-bright, rcp-ndcg-vidore-v3 and rcp-ndcg-trecdl.

All tables are Parquet files and load with datasets or pandas:

from datasets import load_dataset
grades = load_dataset("fabianschmidt-cohere/rcp-ndcg-external-validation", "human_grades", split="train")
import pandas as pd
grades = pd.read_parquet("hf://datasets/fabianschmidt-cohere/rcp-ndcg-external-validation/human_study/grades.parquet")

MANIFEST.tsv lists every file with its size, row count and SHA-256 checksum.

Releases

• Release 2 (this revision, the commit tagged v2; arXiv v2) recomputes the metric side of the human study and of the external-judge comparisons with the release-2 fits of the benchmark datasets, in which the judge's tournament answers are re-read with a corrected parser (see the benchmark cards). The collected data do not change: contests, displayed lists, grades, tie-breaks, reviews, the human majority verdicts, the external judges' answers, and the Study 1 cases and verdicts (Study 1 gains only a flag, below).
  • human_study/contests: d_rcp_shown, rcp_winner_shown, verdict_class and design_cell (the cell of the shown margins) are rescored with the study's scorer on the 311 contests that carry shown margins (7 RCP winners, 4 verdict classes and 7 cells change). The pool columns that selected the contests (ndcg_a/b, rcp_ndcg_a/b, *_winner_pool, is_signflip) keep their study-time values.
  • human_study/contest_documents: theta and pass_prob under the release-2 fit. direct_answer stays the study-time flag (13 rows disagree with the release-2 theta).
  • blind_comparisons/external_judges_cases: d_rcp_ndcg and tercile under the release-2 fit, and d_count_ndcg under release 2 (Count-nDCG does not depend on the fit; its 29 changes outside the side swaps come from the release-2 rule for the rubric answers). In the 32 cases whose RCP side changes, the two sides are swapped (document lists, rerankers, rcp_label_rep1/2, reference preferences), and external_judges_verdicts re-reads rcp_label and choice_side against the new sides (172 of 6,878 verdicts change side; choice_shown is unchanged).
  • blind_comparisons/external_judges_cases has a new column, excluded_under_fit. It is true when a case no longer has, under the release-2 fit, the property it was selected for. In S1 these are 21 of the 480 disagreement cases, where both metrics now favour the same list, and 1 of the 120 control cases, where they now disagree. In S2 they are 9 of the 447 ideal-list cases, and in S3 1 of the 100 minimal pairs, where RCP-nDCG@10 no longer prefers the list it was selected to prefer. The paper's v2 analyses leave these 32 cases out. They are exactly the cases whose sides were swapped. set, kind and stratum keep their selection values, and the tercile in stratum differs from tercile in 39 cases.
  • blind_comparisons/study1_cases has the same new column, under the same rule. A Study 1 case was selected because its shown RCP set had more total RCP gain than the shown qrel set. The flag is true when, under the release-2 fit, the qrel set has at least as much: 3 of the 264 nanomteb_qwen3.5-397b, 0 of the 269 nanomteb_gpt-oss-120b, 13 of the 1,021 bright_qwen3.5-397b, 18 of the 373 bright_qwen3.5-397b_rebuilt and 1 of the 1,042 bright_gpt-oss-120b cases. The paper's v2 Study 1 analyses leave these 35 cases out. No other Study 1 column changes.
  • Every rewritten column keeps its release-1 values in <column>_v1.
• Release 1 is revision 4464fb1a6a0dc2ba2d84cdbb2dd8a070b88accd3 (tag v1).

Conventions

• Suites. nanomteb is NanoBEIR in the paper; the others are bright, vidore and trecdl.
• Identifiers.
  • dataset uses the benchmark's dataset names.
  • query_id and doc_id are the benchmarks' own identifiers, stored as strings.
• Rerankers. Short names as in the paper: RR4-Pro, RR4-Fast, Voyage2.5, Voyage2.5-lt, Qwen3-RR-0.6B, Qwen3-RR-4B, Qwen3-RR-8B, CTXL-RR-1B, CTXL-RR-2B, CTXL-RR-6B, Jina-RR-v3, zerank-1, zerank-1-sm, zerank-2.
• Metric columns. Columns named theta or with the prefix rcp_ refer to the paper's calibrated score and to RCP-nDCG. In human_study/contests, verdict_class = theta means that the verdict supports RCP-nDCG.

human_study/

This is the study with external contracted annotators (Appendix D of the paper). The tables reproduce the paper's counts:

• 311 contests with at least one valid review;
• 46 annotators with a valid review;
• 933 valid reviews;
• 7,080 counted grades on 2,268 distinct documents.

The tables cover the non-test annotator accounts, including the one annotator that quality assurance excluded (flagged).

Table	One row per	Content
contests	contest that received a review (382)	the two rerankers (system_a, system_b), their nDCG@10 and RCP-nDCG@10 over the pool, the winners when the contest was sampled (*_winner_pool), the metric side on the shown documents (*_winner_shown, d_rcp_shown, d_ndcg_shown; @5 over the displayed union with a shared ideal, 0.005 dead band), design_cell, is_signflip, human_verdict (majority of the valid reviewers' picks), human_verdict_old_scorer (an earlier scorer that gave each list its own ideal; the paper's analyses use human_verdict), in_analysis, and the query and instruction as shown to annotators
contest_documents	document shown in a contest	ranks in lists A and B, the qrel, the judge's calibrated score (theta) and gain (pass_prob) under the release-2 fit (rounded to three decimals); theta_v1 and pass_prob_v1 are the values stored when the contest was built; theta_v1 equals the release-1 label up to that rounding wherever the document has one, is_trap for the attention-check document (never scored)
grades	document grade	annotator (hash); facet answers r1_topical, r2_useful (yes/no), r3_answers, r4_complete (yes/no/either); grade (0 Off-topic, 1 Topical, 2 Useful, 3 Answers, 4 Complete; 2.5 and 3.5 come from either); gain (the rubric's gain map); confidence; abstain; carried (pre-filled from the annotator's earlier grade of the same query-document pair); rubric_version; valid_review; counted (enters the paper's analyses)
tiebreaks	pairwise tie-break	winner = a, b or tie, used by the study's scorer
reviews	(contest, annotator) assignment	status, attention-check result (honeypot_passed), QA exclusion, valid_review, the annotator's pick (shared-ideal scorer) and pick_old_scorer
annotators	annotator hash	qualified_datasets (the suites and datasets the annotator was admitted to by subject degree), excluded_by_qa

Further files:

• rubric_versions.json gives, for each version, the rubric shown to annotators: facet questions, help texts, grade labels and gain map.
• guidelines_versions.json gives the guidelines page (HTML) of each version.

blind_comparisons/

These tables hold the parsed cases and verdicts that the paper's analyses use.

• External judges. external_judges_cases and external_judges_verdicts cover the three external judges (GLM-5.3-flash, DeepSeek-4.1-flash, Kimi-K3) on the tournament-order test (Appendices G and H).
  • Sets: S1 (per-query reranker disagreements and agreement controls), S2 (ideal top-10 lists) and S3 (minimal pairs).
  • Each case was sent twice with swapped labels (rep). rcp_label is the label of the RCP side in that replicate, and choice_side is the side the choice favours.
  • reference_prefs_json holds the non-LLM reference preferences of the case.
  • excluded_under_fit marks the cases that the paper's v2 analyses leave out (see Releases).
  • Four of the 6,882 requests have no response.
• Study 1 (Appendix F). study1_cases pairs the qrel set with the tournament-selected set of a query, shown as docs_A and docs_B; label_A and label_B are GT (qrels) or LLM.
  • The set bright_qwen3.5-397b_rebuilt replaces the bright_qwen3.5-397b cases of AoPS, LeetCode and TheoremQA Questions. It removes each query's excluded documents before taking the top k and drops cases whose two sets become identical.
  • study1_external_panel_verdicts gives the three judges' verdicts on every case, each shown twice. verdict_original_labels maps the choice back to the labels of study1_cases, and choice_role names the chosen source.
  • excluded_under_fit in study1_cases marks the cases that the paper's v2 analyses leave out (see Releases).

meta_judges/

These tables hold the full record behind blind_comparisons/: every prompt the three external judges saw and their full answers.

• prompts. One row per prompt, with system_prompt and user_prompt exactly as sent.
  • run is external_judges or study1_panel.
  • rep 2 swaps the two lists and their labels.
  • Prompts join to blind_comparisons/ on these keys: case_id for external_judges; set, dataset and query_id for Study 1.
• responses. The judge's full final answer to each prompt: one row per prompt and judge, the answer that the analyses parsed, joined to prompts on (run, prompt_id). Columns: judge, rep, content and reasoning.
  • Reasoning traces were stored for the tournament-order test only.
  • Parsing content reproduces every verdict in blind_comparisons/.

Anonymization and privacy

• Annotator identities appear only as salted one-way hashes. The salt was discarded, and no table maps hashes back. The hashes equal those of the paper's submitted supplementary data.
• The following are not included: names, email addresses, account, login and session data, all timing information, the annotation workflow log, the accounts of the study team and test accounts, the study team's reference grades, and administrator actions.
• The annotators' optional free-text comments and highlights are not released; the guidelines were screened for email addresses, URLs, phone numbers and account names.
• LLM responses carry public model names. Endpoint addresses and serving identifiers were removed, and error messages have endpoints replaced by [ENDPOINT].

License

• Our contributions are licensed under CC BY-SA 4.0: the human study data, the rubric and guidelines, and the LLM judges' verdicts, reasons, answers and reasoning traces. Please cite the paper (below).
• Benchmark content is not covered by this license and keeps its original terms. This includes queries, instructions, document identifiers, qrels, and the document text inside prompts and LLM-written text: NanoBEIR and its BEIR sources, BRIGHT and ViDoRe v3 (CC BY 4.0), MS MARCO and TREC-DL (non-commercial research use). See LICENSE.

Citation

@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},
}