DOVE: A Large-Scale Multi-Dimensional Predictions Dataset Towards Meaningful LLM Evaluation

Eliya Habba; Ofir Arviv; Itay Itzhak; Yotam Perlitz; Elron Bandel; Leshem Choshen; Michal Shmueli-Scheuer; Gabriel Stanovsky

doi:10.18653/v1/2025.findings-acl.611

DOVE: A Large-Scale Multi-Dimensional Predictions Dataset Towards Meaningful LLM Evaluation

Dataset Of Variation Evaluation — a large collection of LLM answers to the same benchmark questions asked in thousands of slightly different prompt formats

Eliya Habba, Ofir Arviv, Itay Itzhak, Yotam Perlitz, Elron Bandel, Leshem Choshen, Michal Shmueli-Scheuer, Gabriel Stanovsky · Findings of ACL 2025 · 2025

In one sentence

DOVE (Dataset Of Variation Evaluation) is a public dataset of over 250M LLM predictions in which every benchmark instance is perturbed jointly along 5 intent-preserving prompt dimensions, so prompt sensitivity can be measured and mitigated at scale.

Abstract

Recent work found that LLMs are sensitive to a wide range of arbitrary prompt dimensions, including the type of delimiters, answer enumerators, instruction wording, and more. This throws into question popular single-prompt evaluation practices. We present DOVE (Dataset Of Variation Evaluation) a large-scale dataset containing prompt perturbations of various evaluation benchmarks. In contrast to previous work, we examine LLM sensitivity from an holistic perspective, and assess the joint effects of perturbations along various dimensions, resulting in thousands of perturbations per instance. We evaluate several model families against DOVE, leading to several findings, including efficient methods for choosing well-performing prompts, observing that few-shot examples reduce sensitivity, and identifying instances which are inherently hard across all perturbations. DOVE consists of more than 250M prompt perturbations and model outputs, which we make publicly available to spur a community-wide effort toward meaningful, robust, and efficient evaluation. Browse the data, contribute, and more at: https://slab-nlp.github.io/DOVE

Questions this paper answers

is there a big public collection of language model answers to the same questions written many different ways?
what does the DOVE corpus of multi-dimensional prompt perturbations and model outputs cover?
where can I get pre-computed model predictions across thousands of prompt variants instead of generating them myself?
can I download an existing prompt-variation evaluation dataset rather than running my own sweep?: DOVE (Dataset Of Variation Evaluation) is a publicly released dataset of more than 250M LLM prompt perturbations and model outputs, built for studying prompt sensitivity and building more robust evaluation protocols.
Holds for: English multiple-choice benchmarks only (MMLU, MMLU-Pro, ARC, HellaSwag, OpenBookQA, Social IQa, RACE); open-weight Llama 1B/3B/8B, OLMo 7B and Mistral 7B as of the 2025 release; no open-ended generation tasks.; DOVE varies 5 prompt dimensions jointly rather than one dimension at a time, giving over 6.5K perturbations per base instance. The dimensions are enumerator, separator, choices order, instruction phrasing and demonstrations, combined as a Cartesian product.
Holds for: The 5 dimensions were chosen from a survey of prior sensitivity work and are not claimed exhaustive; 78 source datasets with 100 instances each, 7,800 base instances.
which research showed that scoring a model with one fixed prompt gives a misleading benchmark number?
what work should I read first on prompt sensitivity undermining single-prompt benchmark evaluation?
how do I justify to reviewers that a single-template benchmark score is not enough?
should I stop trusting leaderboard numbers that come from one prompt template per benchmark?: DOVE (Dataset Of Variation Evaluation) is a publicly released dataset of more than 250M LLM prompt perturbations and model outputs, built for studying prompt sensitivity and building more robust evaluation protocols.
Holds for: English multiple-choice benchmarks only (MMLU, MMLU-Pro, ARC, HellaSwag, OpenBookQA, Social IQa, RACE); open-weight Llama 1B/3B/8B, OLMo 7B and Mistral 7B as of the 2025 release; no open-ended generation tasks.; Prompt sensitivity does not wash out at scale: OLMo's accuracy on HellaSwag ranges from 1% to 99% depending on which intent-preserving prompt is used.
Holds for: Accuracy computed over 100 instances per dataset per prompt; multiple-choice items scored by semantic similarity matching.; A benchmark's original prompt is often not representative of average performance across intent-preserving prompts. Mistral's accuracy with original instructions exceeds its mean across prompts by more than 1 standard deviation on 35 of 57 MMLU domain tasks.
Holds for: Divergence measured across few-shot prompts, with divergence defined as exceeding 1 standard deviation; MMLU domains, Mistral-7B; full per-model results in Appendix B.3.
how much can a model's score move if you only reword the question or shuffle the answer options?
what is the magnitude of accuracy variance across intent-preserving prompt perturbations on multiple-choice benchmarks?
how do I find out how wide the accuracy range for my model is across formatting variants of the same task?
if I evaluate at a much larger scale, will prompt sensitivity average out?: Prompt sensitivity does not wash out at scale: OLMo's accuracy on HellaSwag ranges from 1% to 99% depending on which intent-preserving prompt is used.
Holds for: Accuracy computed over 100 instances per dataset per prompt; multiple-choice items scored by semantic similarity matching.; Each individual prompt dimension moves accuracy even after averaging over all the others: for Mistral, different instruction paraphrases alone produce an 8% accuracy difference.
Holds for: Marginalized accuracy over the 5 DOVE dimensions, computed per domain for Llama 1B/3B/8B, OLMo 7B and Mistral 7B; 13 instruction phrasings.
do different language models do best with different answer labels like A/B/C, roman numerals or greek letters?
are optimal enumerator choices model-specific or is there a universally best multiple-choice format?
how do I pick answer-option labels for a multiple-choice task across several models at once?
can I reuse the prompt format that worked well for one model on a different model?: Models have distinct prompt preferences rather than a shared best format: greek numerals are OLMoE's best-performing enumerator of the 6 tested but only Mistral's third best, behind capital and lateen numerals.
Holds for: 6 enumerator values, accuracy marginalized over all other dimensions and averaged across the dataset; established for the evaluated open-weight models only.
is it better to pick the single best prompt you tried, or to choose each formatting choice separately?
does dimension-wise marginal selection beat argmax over observed prompts for prompt-format optimization?
how do I choose a prompt format for a multiple-choice task with as few trial runs as possible?
I can only afford a few hundred prompt trials, how should I use them to settle on a format?: Choosing the best value for each of the 5 prompt dimensions independently, marginalizing the others, matches linear regression on observed samples and beats picking the single best observed prompt.
Holds for: Zero-shot selection simulated on DOVE against the best prompt over the full dataset, mean and standard deviation over 10 random seeds; 4 strategies including a random baseline.; Selecting the best observed prompt only becomes a reliable strategy after tens of millions of observed samples, while dimension-wise selection works with relatively small sample sizes.
Holds for: Zero-shot setting, gap from the best prompt in DOVE as a function of sample count, for Llama 1B/3B/8B, OLMo 7B and Mistral 7B; AUC comparison across models in Appendix B.4.
does showing a few worked examples in the prompt make a model's score less dependent on wording?
does 5-shot prompting reduce accuracy variance across prompt perturbations relative to 0-shot?
how do I make benchmark scores more stable across formatting changes?
if I switch to few-shot prompting, do I still need to worry about prompt sensitivity?: Five-shot prompting consistently narrows the spread of accuracy across prompt perturbations compared to zero-shot, but does not eliminate it: the accuracy range still exceeds 20% on every dataset shown.
Holds for: Subset of DOVE domains, each point an accuracy over 100 instances; the effect is minimal for Social IQa and the MMLU-Pro legal domain.
how can you tell which test questions a model gets wrong no matter how you ask them?
how are success-rate distributions over prompt perturbations used to identify inherently hard instances?
how do I separate genuinely difficult benchmark items from ones my model fails only for formatting reasons?
should I spend effort on prompt engineering for the items my model keeps failing, or are they hopeless?: Success-rate distributions over the more than 6.5K perturbations per instance identify instances a model answers wrongly under every perturbation. Such instances give an operational definition of inherently hard items that no prompt selection can fix.
Holds for: Success rate is the percentage of a sample's perturbations answered correctly; measured per model on DOVE's multiple-choice instances; at both extremes models are also least prompt-sensitive.
is the prompt that ships with a benchmark like MMLU typical of how a model does on average?
how far does accuracy under a benchmark's original template diverge from the mean over intent-preserving prompts?
how do I check whether the default template I am evaluating with is flattering my model?
should I report the default MMLU prompt score or an average over prompt variants?: A benchmark's original prompt is often not representative of average performance across intent-preserving prompts. Mistral's accuracy with original instructions exceeds its mean across prompts by more than 1 standard deviation on 35 of 57 MMLU domain tasks.
Holds for: Divergence measured across few-shot prompts, with divergence defined as exceeding 1 standard deviation; MMLU domains, Mistral-7B; full per-model results in Appendix B.3.
how much compute and money does it take to test one model on thousands of reworded versions of every question?
what GPU-hour and cloud budget was needed to generate over 250M prompt perturbation outputs?
how do I budget for an evaluation that sweeps thousands of prompt variants per instance?
can I afford to run a full multi-prompt sensitivity evaluation on my own hardware?: Building DOVE required roughly 5,000 GPU hours on NVIDIA A100 80GB GPUs, with Mistral-7B alone taking 1,189 GPU hours, and would cost upwards of $25K on cloud services such as AWS.
Holds for: Generation with vLLM at approximately 60M instances per model for 5 open-weight models; closed API models were excluded partly for this reason.
which parts of a prompt get changed in a large prompt-variation dataset, and how many versions of each question are there?
which perturbation dimensions does DOVE vary, and what is the per-instance perturbation count from their Cartesian product?
how do I build a prompt sweep that varies separators, option order and instruction phrasing together instead of one at a time?
does a prompt-variation dataset cover the formatting axes I actually care about, like demonstrations and option order?: DOVE varies 5 prompt dimensions jointly rather than one dimension at a time, giving over 6.5K perturbations per base instance. The dimensions are enumerator, separator, choices order, instruction phrasing and demonstrations, combined as a Cartesian product.
Holds for: The 5 dimensions were chosen from a survey of prior sensitivity work and are not claimed exhaustive; 78 source datasets with 100 instances each, 7,800 base instances.
can other research groups contribute their own model outputs to a shared prompt-variation dataset?
is the DOVE resource maintained as a living benchmark with a standardized schema for contributed evaluation data?
how do I add predictions from my own model to a shared prompt-sensitivity dataset without re-running everything?
if I contribute my evaluation runs, will they be usable alongside the existing prompt perturbation data?: DOVE is maintained as an extensible community resource with a standardized schema for multi-dimensional evaluation data, so that model predictions from other groups can be converted in rather than re-run.
Holds for: As of the 2025 Findings of ACL release; contributions arrive via HuggingFace pull requests or email; coverage at release is English-centric and limited in model diversity.

Claims and scope

DOVE (Dataset Of Variation Evaluation) is a publicly released dataset of more than 250M LLM prompt perturbations and model outputs, built for studying prompt sensitivity and building more robust evaluation protocols. (Section 1)
Scope: English multiple-choice benchmarks only (MMLU, MMLU-Pro, ARC, HellaSwag, OpenBookQA, Social IQa, RACE); open-weight Llama 1B/3B/8B, OLMo 7B and Mistral 7B as of the 2025 release; no open-ended generation tasks.
DOVE varies 5 prompt dimensions jointly rather than one dimension at a time, giving over 6.5K perturbations per base instance. The dimensions are enumerator, separator, choices order, instruction phrasing and demonstrations, combined as a Cartesian product. (Table 1)
Scope: The 5 dimensions were chosen from a survey of prior sensitivity work and are not claimed exhaustive; 78 source datasets with 100 instances each, 7,800 base instances.
Building DOVE required roughly 5,000 GPU hours on NVIDIA A100 80GB GPUs, with Mistral-7B alone taking 1,189 GPU hours, and would cost upwards of $25K on cloud services such as AWS. (Section 4.1)
Scope: Generation with vLLM at approximately 60M instances per model for 5 open-weight models; closed API models were excluded partly for this reason.
Prompt sensitivity does not wash out at scale: OLMo's accuracy on HellaSwag ranges from 1% to 99% depending on which intent-preserving prompt is used. (Figure 3)
Scope: Accuracy computed over 100 instances per dataset per prompt; multiple-choice items scored by semantic similarity matching.
Each individual prompt dimension moves accuracy even after averaging over all the others: for Mistral, different instruction paraphrases alone produce an 8% accuracy difference. (Figure 4)
Scope: Marginalized accuracy over the 5 DOVE dimensions, computed per domain for Llama 1B/3B/8B, OLMo 7B and Mistral 7B; 13 instruction phrasings.
Models have distinct prompt preferences rather than a shared best format: greek numerals are OLMoE's best-performing enumerator of the 6 tested but only Mistral's third best, behind capital and lateen numerals. (Figure 4)
Scope: 6 enumerator values, accuracy marginalized over all other dimensions and averaged across the dataset; established for the evaluated open-weight models only.
A benchmark's original prompt is often not representative of average performance across intent-preserving prompts. Mistral's accuracy with original instructions exceeds its mean across prompts by more than 1 standard deviation on 35 of 57 MMLU domain tasks. (Figure 5)
Scope: Divergence measured across few-shot prompts, with divergence defined as exceeding 1 standard deviation; MMLU domains, Mistral-7B; full per-model results in Appendix B.3.
Choosing the best value for each of the 5 prompt dimensions independently, marginalizing the others, matches linear regression on observed samples and beats picking the single best observed prompt. (Figure 6)
Scope: Zero-shot selection simulated on DOVE against the best prompt over the full dataset, mean and standard deviation over 10 random seeds; 4 strategies including a random baseline.
Selecting the best observed prompt only becomes a reliable strategy after tens of millions of observed samples, while dimension-wise selection works with relatively small sample sizes. (Figure 7)
Scope: Zero-shot setting, gap from the best prompt in DOVE as a function of sample count, for Llama 1B/3B/8B, OLMo 7B and Mistral 7B; AUC comparison across models in Appendix B.4.
Five-shot prompting consistently narrows the spread of accuracy across prompt perturbations compared to zero-shot, but does not eliminate it: the accuracy range still exceeds 20% on every dataset shown. (Figure 8)
Scope: Subset of DOVE domains, each point an accuracy over 100 instances; the effect is minimal for Social IQa and the MMLU-Pro legal domain.
Success-rate distributions over the more than 6.5K perturbations per instance identify instances a model answers wrongly under every perturbation. Such instances give an operational definition of inherently hard items that no prompt selection can fix. (Figure 9)
Scope: Success rate is the percentage of a sample's perturbations answered correctly; measured per model on DOVE's multiple-choice instances; at both extremes models are also least prompt-sensitive.
DOVE is maintained as an extensible community resource with a standardized schema for multi-dimensional evaluation data, so that model predictions from other groups can be converted in rather than re-run. (Section 6)
Scope: As of the 2025 Findings of ACL release; contributions arrive via HuggingFace pull requests or email; coverage at release is English-centric and limited in model diversity.

Common misreadings

DOVE is not a new benchmark of questions: it is a dataset of model predictions, log probabilities and scores built on top of existing benchmarks such as MMLU, ARC and RACE.
The finding that few-shot demonstrations reduce sensitivity does not mean five-shot evaluation is stable: the accuracy range across perturbations remains above 20% on every dataset measured, and the effect is minimal for Social IQa and the MMLU-Pro legal domain.
The 1%-to-99% HellaSwag range is a range across individual intent-preserving prompts, not a confidence interval on a single score; each endpoint is an accuracy obtained with some legitimate prompt format.
DOVE does not show that prompt sensitivity is confined to small open-weight models; only open-weight models up to 8B were run, because API models alter prompts in undisclosed ways and costs were infeasible.
"Independent dimension-wise selection works best" is a statement about sample efficiency under a fixed budget, not a claim that prompt dimensions do not interact; with tens of millions of samples the best-observed-prompt strategy becomes reliable too.

Terminology in this paper

Intent-preserving prompts: Two prompts designed to convey the same underlying meaning to a model despite differences in phrasing or structure, e.g. adding "Answer the following question:" before the same multiple-choice item.
Prompt dimension: A set of interchangeable prompt choices whose values all preserve the prompt's intent, such as the answer enumerator (capitals, lowercase, numbers, roman numerals, keyboard symbols, greek letters).
Prompt linearization: The deterministic function that maps an underlying question plus one chosen value per prompt dimension into a single concrete prompt string given to an LLM.
Marginalized accuracy: The accuracy of a model for one fixed value of a single prompt dimension, averaged over all combinations of values of the remaining dimensions.
Divergence score: The number of standard deviations by which a model's accuracy using a benchmark's original prompt deviates from its mean accuracy across all prompt perturbations, with divergence counted when it exceeds 1.
Success rate: For one benchmark instance and one model, the percentage of that instance's prompt perturbations for which the model outputs the correct answer.
Inherently hard instance: A benchmark instance a model answers incorrectly under every prompt perturbation tested, so no prompt selection can recover it.

How to cite

@inproceedings{habba-etal-2025-dove,
    title = "{DOVE}: A Large-Scale Multi-Dimensional Predictions Dataset Towards Meaningful {LLM} Evaluation",
    author = "Habba, Eliya  and
      Arviv, Ofir  and
      Itzhak, Itay  and
      Perlitz, Yotam  and
      Bandel, Elron  and
      Choshen, Leshem  and
      Shmueli-Scheuer, Michal  and
      Stanovsky, Gabriel",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.findings-acl.611/",
    doi = "10.18653/v1/2025.findings-acl.611",
    pages = "11744--11763",
    ISBN = "979-8-89176-256-5",
    abstract = "Recent work found that LLMs are sensitive to a wide range of arbitrary prompt dimensions, including the type of delimiters, answer enumerators, instruction wording, and more. This throws into question popular single-prompt evaluation practices. We present DOVE (Dataset Of Variation Evaluation) a large-scale dataset containing prompt perturbations of various evaluation benchmarks. In contrast to previous work, we examine LLM sensitivity from an holistic perspective, and assess the joint effects of perturbations along various dimensions, resulting in thousands of perturbations per instance. We evaluate several model families against DOVE, leading to several findings, including efficient methods for choosing well-performing prompts, observing that few-shot examples reduce sensitivity, and identifying instances which are inherently hard across all perturbations. DOVE consists of more than 250M prompt perturbations and model outputs, which we make publicly available to spur a community-wide effort toward meaningful, robust, and efficient evaluation. Browse the data, contribute, and more at: https://slab-nlp.github.io/DOVE"
}

References

See the full reference list in the paper.