Skip to content

Repository files navigation

The Multivac: Blind Peer Matrix Evaluation Dataset

License: MIT

Overview

This dataset contains 27,540 LLM judgments from a blind peer matrix evaluation of 55 frontier language models across 286 evaluations and 198 unique questions in 9 category pools. Of the 27,540: 23,356 parsed successfully and 22,252 carry a usable score (the analysis set); 2,781 are intentional self-exclusions (the matrix diagonal) and 1,403 are judge failures (parse/API errors). To our knowledge, this is the largest publicly available multi-judge LLM evaluation dataset with full provenance. See scripts/count_reconciliation.py for the exact breakdown.

Paper: The Multivac: Blind Peer Matrix Evaluation of Frontier Language Models

Dataset Structure

multivac-evaluation/
├── README.md                          # This file
├── LICENSE                            # MIT License
├── DATASHEET.md                       # NeurIPS Datasheet for Datasets
├── CITATION.cff                       # Citation metadata
├── evaluation_framework/
│   ├── multivac.py                    # Core evaluation engine
│   ├── extract_multivac_data.py       # Data extraction pipeline
│   ├── statistical_analysis.py        # Statistical tests
│   ├── questions.py                   # Wave 1 question bank
│   └── questions_wave2_15032026.py    # Wave 2 question bank
├── data/
│   ├── peer_matrix/                   # 286 EVAL-* folders
│   └── head_to_head/                  # H2H batch folders
└── paper_tables/                      # Pre-computed tables from the paper

Key Statistics

Metric Value
Peer matrix evaluations 286
Unique questions 198
Total judgments 27,540
Parsed judgments 23,356
Usable-scored (analysis set) 22,252
Self-excluded (diagonal) 2,781
Judge failures (parse/API) 1,403
Unique models 55
Vendor families 17
Category pools 9
H2H questions (recorded) 185
Date range Feb–Apr 2026

Key Findings

  1. No category has a statistically separated winner — in all nine pools the top model beats second place in only 50–58% of matched comparisons, with the 95% interval spanning 50%. Four distinct models lead the six primary pools by point estimate, but none separates from its runner-up (scripts/make_finding1.py)
  2. A same-vendor scoring premium survives correction for only 2 of 8 families — under a within-response fixed-effects model with standard errors clustered two-way by judge and question, Anthropic (+0.41) is robust (worst-case leave-one-judge-out p < 10⁻⁴) and MiniMax (+0.40) is significant under the primary specification (p = 0.001) but judge-dependent (its two-way significance does not survive leave-one-judge-out, worst-case p ≈ 0.02); Qwen (+0.56) does not survive the two-way clustering and is not counted. The large naive estimates — including Mistral −1.02 and Google −0.59 — are artifacts of judge leniency and respondent quality, not a same-vendor premium, and do not survive controls. The two surviving premiums are an upper bound on favoritism, not a measurement of it: the estimator cannot separate preferential scoring from a sibling judge parsing same-distribution output more accurately. See paper_tables/FOUR_CELL_DECOMPOSITION_FINDINGS.md and WITHIN_RESPONSE_FINDINGS.md. (Supersedes the earlier "significant bias in all families" claim.)
  3. The "overall" champion is an artifact of the aggregation — naive mean, leniency-adjusted Bradley–Terry, and exposure-restricted BT each crown a different model. Among broad-participation generalists GPT-5.4 holds rank 1 in 99% of question-clustered resamples, with a 54% per-comparison edge over the runner-up (scripts/make_finding1.py). (An earlier bootstrap on the naive-mean leaderboard — where rank 1 is Grok 4.1 Fast — gave p = 0.266/0.073/0.071 vs ranks 2–4; that aggregate is leniency-confounded and is superseded by the BT rankings.)
  4. Judge disagreement is category-dependent — code σ=1.27 vs meta-alignment σ=0.63 (ratio 2.01×; scripts/category_disagreement.py)
  5. Overall inter-annotator agreement: Krippendorff's α = 0.618 (evaluation_framework/statistical_analysis.py)

Intended Uses

  • Model selection: Category-specific rankings for task-appropriate model choice
  • DPO/RLHF training data: 90 scored judgments per evaluation, convertible to preference pairs, at <$0.01/sample
  • Judge bias research: Analysis of systematic judge behavior and family bias
  • Evaluation methodology research: Comparing single-judge vs multi-judge approaches
  • Safety monitoring: Tracking alignment behavior across model versions

Limitations

  • All judgments are from LLMs, not humans. Shared biases across models are not captured.
  • Question selection reflects one author's judgment of what constitutes a good evaluation prompt.
  • Model participation is non-uniform: core-pool models appear in 163–238 evaluations, focused-batch models in as few as 1.
  • See §6 of the paper for full limitations discussion.

Citation

@article{darji2026multivac,
  title={The Multivac: Blind Peer Matrix Evaluation of Frontier Language Models},
  author={Darji, Yash},
  journal={arXiv preprint arXiv:TODO},
  year={2026}
}

License

MIT License. See LICENSE for details.

Contact

About

No description, website, or topics provided.

Resources

Stars

6 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages