Hugging Face's logo

Datasets:
JevArena-mm
/
jevarena-MM_image

Tasks:
Visual Question Answering
Image Classification
Modalities:
Image
Languages:
English
Tags:
benchmark
calibration
multimodal
image
License:
Dataset card Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

These are the image files of public research datasets, re-hosted so that JevArena-MM can be run. Each file stays under its source's license (table below); several sources allow non-commercial research use only. By requesting access you agree to use the files only to evaluate models or for other non-commercial research, to follow each source's license and terms, not to redistribute them, and to delete any file whose owner asks for its removal.

Log in or Sign Up to review the conditions and access this dataset content.

JevArena-MM (image)

The image tasks of JevArena-MM, a benchmark of JevArena decision requests whose state carries images. This dataset holds the items, the code that builds, runs and scores them, the result tables, and the files the items point to.

  • A file an item references as <folder>/<path> is at <folder>/<path> here, so a download of this repository is a media root (JEVARENA_MM_MEDIA). Items record each file's sha256.
  • Every source image is the source's original file, unchanged, at <source>/<path in the release>. Images the image team constructed (rendered clocks, plots and boards, crops, drawn boxes, masked or composed images) are under jevarena_img/<leaf>.

Status: unreleased, all tasks draft. 47 tasks over 40 of the 42 image leaves (23,000 items), plus 7 standard benchmarks in img.std (ScienceQA, A-OKVQA, OK-VQA, VizWiz-VQA, GQA, MMMU, NaturalBench; 3,455 items), written to the JevArena-MM proposal of 2026-10-09.

Download

Accept the terms on this page, log in with hf auth login (or set HF_TOKEN), then in a clone of this repository:

pip install -e '.[image]'
export JEVARENA_MM_MEDIA=/path/to/media
jevarena-mm fetch --leaf img.        # the folders the items need, from this dataset
jevarena-mm verify

One folder without the code:

hf download JevArena-mm/jevarena-MM_image --repo-type dataset --include 'mmmu/*' --local-dir /path/to/media

Hosted folders

Folder Source License Files are Files Size
agentrewardbench AgentRewardBench, test and dev (Lù et al., 2025) dataset terms of use (research) original 1,000 0.19 GB
aokvqa A-OKVQA, validation split, multiple-choice setting (Schwenk et al., 2022) Apache-2.0 (annotations); COCO images original 499 0.03 GB
atari_head Atari-HEAD, Breakout and Ms. Pac-Man trials (Zhang et al., 2019) CC BY 4.0 original 500 0.00 GB
blink BLINK, val (Fu et al., 2024) Apache-2.0 original 1,020 0.14 GB
bongard_openworld Bongard-OpenWorld, test (Wu et al., 2024) CC BY-NC-SA 4.0; images crawled from the web original 2,797 0.93 GB
charxiv CharXiv, val (Wang et al., 2024) CC BY-SA 4.0 (questions); charts by arXiv authors original 97 0.01 GB
cord_v2 CORD v2, train/validation/test (Park et al., 2019) CC BY 4.0 original 500 1.15 GB
countbench CountBench (Paiss et al., 2023) unspecified; web images original 469 0.02 GB
cub_200_2011 CUB-200-2011, test split (Wah et al., 2011); CUB-200-2011, test split, attribute labels (Wah et al., 2011) CC BY original 958 0.09 GB
eb_alfred EmbodiedBench EB-ALFRED trajectories, single step (Yang et al., 2025) not stated on the dataset card original 500 0.12 GB
esnlive e-SNLI-VE, test split (Kayser et al., 2021) unspecified on HF (Flickr30k images, SNLI-VE labels) original 500 0.07 GB
figstep FigStep SafeBench (Gong et al., 2025) MIT original 148 0.01 GB
genai_bench GenAI-Bench, train (Li et al., 2024) Apache-2.0 original 387 0.23 GB
glitchbench GlitchBench, validation (Taesiri et al., 2024) MIT (card); screenshots from Reddit original 249 0.30 GB
gqa GQA, testdev balanced split (Hudson & Manning, 2019) CC BY 4.0 original 252 0.04 GB
hallusionbench HallusionBench, image split (Guan et al., 2024) BSD-3-Clause original 320 0.12 GB
hateful_memes Hateful Memes, test_seen (Kiela et al., 2020) Hateful Memes Dataset License (non-commercial research) original 500 0.18 GB
infographicvqa InfographicVQA, validation (Mathew et al., 2022) Apache-2.0 (HF mirror); InfographicVQA terms original 494 0.20 GB
jevarena_img/img.code_render_match Program-generated matplotlib scripts and renders (image team) CC BY 4.0 (generated) constructed by the image team 500 0.01 GB
jevarena_img/img.game_board_twin Lichess puzzles, rendered boards (Lichess, 2024) CC0 constructed by the image team 500 0.04 GB
jevarena_img/img.know_clock_calendar Program-rendered clocks and calendars (image team) CC BY 4.0 (generated) constructed by the image team 1,000 0.02 GB
jevarena_img/img.know_perception CountBench (Paiss et al., 2023) unspecified; web images constructed by the image team 10 0.00 GB
jevarena_img/img.know_visual_rules Bongard-OpenWorld, test (Wu et al., 2024) CC BY-NC-SA 4.0; images crawled from the web constructed by the image team 3 0.00 GB
jevarena_img/img.pro_damage xBD / xView2, test (Gupta et al., 2019) CC BY-NC-SA 4.0 constructed by the image team 1,000 0.03 GB
jevarena_img/img.route_abstain Masked copies of MMMU-Pro items (image team) Apache-2.0 (MMMU-Pro) constructed by the image team 500 0.10 GB
jevarena_img/img.route_model_correctness MMMU-Pro items with Qwen2.5-VL-3B-Instruct outcomes (image team) Apache-2.0 (MMMU-Pro) constructed by the image team 517 0.15 GB
jevarena_img/img.safety_injection FigStep SafeBench (Gong et al., 2025) MIT constructed by the image team 147 0.00 GB
jevarena_img/img.safety_synthetic GenImage Midjourney subset, val (Zhu et al., 2023) GenImage terms (research); ImageNet images constructed by the image team 500 0.04 GB
jevarena_img/img.social_art_emotion ArtEmis v2 contrastive, WikiArt images (Mohamed et al., 2022) ArtEmis terms (research use) constructed by the image team 1 0.00 GB
jevarena_img/img.step_grounding AndroidControl, test (Li et al., 2024) Apache-2.0 constructed by the image team 500 0.24 GB
jevarena_img/img.step_gui_action Multimodal-Mind2Web, test_website (Deng et al., 2023) OpenRAIL constructed by the image team 249 0.19 GB
koniq10k KonIQ-10k, test (Hosu et al., 2020) KonIQ-10k research terms original 500 0.26 GB
mathvista MathVista, testmini (Lu et al., 2024) CC BY-SA 4.0 original 495 0.05 GB
mmmu MMMU, validation split, multiple-choice questions (Yue et al., 2024) Apache-2.0 (release); images from the sources it cites original 550 0.19 GB
mmmu_pro MMMU-Pro, standard 10 options, test (Yue et al., 2024) Apache-2.0 original 518 0.20 GB
mvtec_ad MVTec AD, test (Bergmann et al., 2019) CC BY-NC-SA 4.0 original 500 0.49 GB
naturalbench NaturalBench (Li et al., 2024) Apache-2.0 original 498 0.15 GB
newyorker_caption_ranking New Yorker caption ranking, test (Zhang et al., 2024) CC BY-NC 4.0; cartoons copyrighted original 47 0.02 GB
nlvr2 NLVR2, balanced public test (Suhr et al., 2019) CC BY 4.0 annotations; web images original 1,000 0.10 GB
okvqa OK-VQA, val2014 split (Marino et al., 2019) CC BY 4.0 (annotations); COCO images original 500 0.08 GB
phishpedia Phishpedia benchmark samples (Lin et al., 2021) research use (unspecified) original 500 0.19 GB
resisc45 RESISC45, test (Cheng et al., 2017) unknown (HF card) original 500 0.01 GB
rvl_cdip RVL-CDIP, test (Harley et al., 2015) research use (IIT-CDIP) original 500 0.06 GB
scienceqa ScienceQA, test split, questions with an image (Lu et al., 2022) CC BY-NC-SA 4.0 original 488 0.03 GB
seetrue SeeTRUE, test (Yarom et al., 2023) CC BY 4.0, test-only use original 500 0.22 GB
slake SLAKE, test (Liu et al., 2021) CC BY 4.0 original 96 0.01 GB
stanford_online_products Stanford Online Products, Ebay_test (Oh Song et al., 2016) research use (unspecified) original 1,000 0.02 GB
tabmwp TabMWP, test (Lu et al., 2023) CC BY-SA 4.0 original 500 0.01 GB
unsafebench UnsafeBench, test (Qu et al., 2024) see dataset card (gated) original 500 0.04 GB
videogameqa_bench VideoGameQA-Bench, test (Taesiri et al., 2025) CC BY 4.0 original 500 0.27 GB
vispr VISPR, test 2017 (Orekondy et al., 2017) CC BY-NC 4.0 original 500 1.13 GB
vizwiz VizWiz-QualityIssues, val (Chiu et al., 2020); VizWiz-VQA, val (Gurari et al., 2018); VizWiz-VQA, validation split (Gurari et al., 2018) CC BY 4.0 original 1,500 0.69 GB
vl_rewardbench VL-RewardBench, test (Li et al., 2025) MIT original 250 0.01 GB
wikiart ArtEmis v2 contrastive, WikiArt images (Mohamed et al., 2022) ArtEmis terms (research use) original 499 0.16 GB
winoground Winoground, test (Thrush et al., 2022) Winoground research license (click-through) original 500 0.23 GB

29,558 files, 9.3 GB in 55 folders.

Results (unreleased, draft items)

These tables were run on the image team's pipeline-built copies (resized and re-saved), before the items switched to original files (2026-10-11). Responses whose model input is byte-for-byte unchanged were carried over to the new items; the others need a re-run (results/carry_over.json). The tables are re-scored after that re-run.

All 17 models were scored on the 23,000 image items (40 leaves). avg@1 is accuracy: an item's questions are averaged, a leaf is the mean of its items, and the overall score is the unweighted mean of the leaves. Brier and ECE (10 equal-mass bins) measure calibration; lower is better. Brier is scored against the annotators' label distribution where the source has one (KonIQ-10k, ArtEmis, New Yorker), otherwise against the gold label; "Brier (one-hot)" uses the gold label everywhere. Model specs are listed in results/summary/unreleased/README.md.

Generated by jevarena-mm compare. avg@1 (accuracy); coverage in brackets where a model was not run on every item; the best value of each column in bold.

Overall

# model avg@1 maj@1 Brier Brier (one-hot) ECE coverage
1 OneJev-27B 0.678 0.678 0.388 0.417 0.077 1.000
2 OneJev-27B-GGUF-Q8_0 0.678 0.678 0.388 0.417 0.079 1.000
3 Qwen3.8-27B-FP8 0.647 0.647 0.439 0.467 0.113 1.000
4 OneJev-9B 0.645 0.645 0.426 0.455 0.085 1.000
5 Qwen3.5-35B-A3B-GPTQ-Int4 0.645 0.645 0.461 0.490 0.147 1.000
6 OneJev-4B 0.624 0.624 0.444 0.473 0.083 1.000
7 Qwen3.5-9B 0.606 0.606 0.494 0.523 0.150 1.000
8 diffusiongemma-26B-A4B-it 0.602 0.602 0.647 0.677 0.297 0.978
9 Qwen3-Omni-30B-A3B-Thinking 0.575 0.575 0.665 0.695 0.289 1.000
10 Qwen3.5-4B 0.572 0.572 0.521 0.551 0.137 1.000
11 Qwen2.5-Omni-7B 0.553 0.553 0.569 0.594 0.169 1.000
12 Qwen2.5-Omni-7B-GPTQ-Int4 0.546 0.546 0.560 0.586 0.153 1.000
13 Qwen2.5-Omni-7B-AWQ 0.544 0.544 0.571 0.596 0.169 1.000
14 Qwen3-Omni-30B-A3B-Instruct 0.544 0.544 0.671 0.701 0.277 1.000
15 OneJev-0.8B 0.519 0.519 0.548 0.578 0.099 1.000
16 Qwen2.5-Omni-3B 0.495 0.495 0.584 0.612 0.136 1.000
17 Qwen3.5-0.8B 0.387 0.387 0.695 0.728 0.211 1.000

Overall is the mean of the taxonomy leaves; the standard benchmarks (img.std) are reported in their own tables.

By domain

Each domain is the unweighted mean of its leaves.

model basic batch code embodied game judge know pro ret route safety social std step struct
OneJev-27B 0.654 0.705 0.584 0.594 0.423 0.809 0.641 0.716 0.915 0.600 0.751 0.293 0.859 0.732 0.889
OneJev-27B-GGUF-Q8_0 0.652 0.702 0.588 0.592 0.422 0.811 0.643 0.715 0.914 0.600 0.748 0.301 0.730 0.887
Qwen3.8-27B-FP8 0.635 0.632 0.554 0.590 0.412 0.813 0.565 0.666 0.873 0.560 0.717 0.374 0.667 0.861
OneJev-9B 0.653 0.746 0.424 0.532 0.414 0.786 0.587 0.676 0.867 0.578 0.730 0.293 0.661 0.839
Qwen3.5-35B-A3B-GPTQ-Int4 0.666 0.663 0.440 0.504 0.435 0.808 0.560 0.692 0.873 0.586 0.741 0.259 0.605 0.855
OneJev-4B 0.610 0.723 0.418 0.592 0.384 0.762 0.564 0.657 0.847 0.590 0.712 0.263 0.649 0.771
Qwen3.5-9B 0.629 0.696 0.416 0.522 0.402 0.739 0.469 0.629 0.888 0.570 0.711 0.235 0.589 0.815
diffusiongemma-26B-A4B-it 0.559 0.523 0.470 0.552 0.384 0.783 0.503 0.670 0.854 0.535 0.708 (0.78) 0.309 0.549 0.796
Qwen3-Omni-30B-A3B-Thinking 0.559 0.713 0.318 0.494 0.414 0.730 0.462 0.590 0.859 0.580 0.661 0.210 0.502 0.732
Qwen3.5-4B 0.531 0.528 0.402 0.552 0.395 0.708 0.458 0.580 0.870 0.494 0.694 0.419 0.504 0.721
Qwen2.5-Omni-7B 0.573 0.744 0.230 0.508 0.417 0.666 0.451 0.501 0.833 0.481 0.698 0.343 0.465 0.680
Qwen2.5-Omni-7B-GPTQ-Int4 0.551 0.712 0.220 0.522 0.420 0.688 0.432 0.502 0.842 0.507 0.675 0.329 0.433 0.644
Qwen2.5-Omni-7B-AWQ 0.574 0.721 0.258 0.510 0.413 0.661 0.437 0.504 0.804 0.492 0.659 0.361 0.441 0.667
Qwen3-Omni-30B-A3B-Instruct 0.583 0.511 0.300 0.514 0.353 0.693 0.465 0.529 0.855 0.468 0.582 0.315 0.506 0.762
OneJev-0.8B 0.489 0.659 0.254 0.544 0.374 0.689 0.413 0.534 0.804 0.493 0.608 0.339 0.471 0.500
Qwen2.5-Omni-3B 0.480 0.635 0.232 0.528 0.370 0.566 0.395 0.480 0.802 0.466 0.638 0.342 0.364 0.535
Qwen3.5-0.8B 0.342 0.456 0.230 0.464 0.293 0.509 0.322 0.331 0.650 0.387 0.525 0.136 0.239 0.406

By leaf

One table per domain: the domain mean, then each leaf.

img.basic

model mean custom_categories fine_grained visual_entailment
OneJev-27B 0.654 0.697 0.412 0.853
OneJev-27B-GGUF-Q8_0 0.652 0.695 0.406 0.855
Qwen3.8-27B-FP8 0.635 0.691 0.366 0.847
OneJev-9B 0.653 0.681 0.464 0.813
Qwen3.5-35B-A3B-GPTQ-Int4 0.666 0.661 0.478 0.858
OneJev-4B 0.610 0.661 0.370 0.799
Qwen3.5-9B 0.629 0.660 0.404 0.823
diffusiongemma-26B-A4B-it 0.559 0.640 0.236 0.800
Qwen3-Omni-30B-A3B-Thinking 0.559 0.568 0.358 0.752
Qwen3.5-4B 0.531 0.617 0.196 0.779
Qwen2.5-Omni-7B 0.573 0.574 0.362 0.783
Qwen2.5-Omni-7B-GPTQ-Int4 0.551 0.571 0.324 0.757
Qwen2.5-Omni-7B-AWQ 0.574 0.566 0.370 0.786
Qwen3-Omni-30B-A3B-Instruct 0.583 0.599 0.430 0.720
OneJev-0.8B 0.489 0.633 0.170 0.665
Qwen2.5-Omni-3B 0.480 0.505 0.290 0.646
Qwen3.5-0.8B 0.342 0.511 0.028 0.486

img.batch

model mean curation
OneJev-27B 0.705 0.705
OneJev-27B-GGUF-Q8_0 0.702 0.702
Qwen3.8-27B-FP8 0.632 0.632
OneJev-9B 0.746 0.746
Qwen3.5-35B-A3B-GPTQ-Int4 0.663 0.663
OneJev-4B 0.723 0.723
Qwen3.5-9B 0.696 0.696
diffusiongemma-26B-A4B-it 0.523 0.523
Qwen3-Omni-30B-A3B-Thinking 0.713 0.713
Qwen3.5-4B 0.528 0.528
Qwen2.5-Omni-7B 0.744 0.744
Qwen2.5-Omni-7B-GPTQ-Int4 0.712 0.712
Qwen2.5-Omni-7B-AWQ 0.721 0.721
Qwen3-Omni-30B-A3B-Instruct 0.511 0.511
OneJev-0.8B 0.659 0.659
Qwen2.5-Omni-3B 0.635 0.635
Qwen3.5-0.8B 0.456 0.456

img.code

model mean render_match
OneJev-27B 0.584 0.584
OneJev-27B-GGUF-Q8_0 0.588 0.588
Qwen3.8-27B-FP8 0.554 0.554
OneJev-9B 0.424 0.424
Qwen3.5-35B-A3B-GPTQ-Int4 0.440 0.440
OneJev-4B 0.418 0.418
Qwen3.5-9B 0.416 0.416
diffusiongemma-26B-A4B-it 0.470 0.470
Qwen3-Omni-30B-A3B-Thinking 0.318 0.318
Qwen3.5-4B 0.402 0.402
Qwen2.5-Omni-7B 0.230 0.230
Qwen2.5-Omni-7B-GPTQ-Int4 0.220 0.220
Qwen2.5-Omni-7B-AWQ 0.258 0.258
Qwen3-Omni-30B-A3B-Instruct 0.300 0.300
OneJev-0.8B 0.254 0.254
Qwen2.5-Omni-3B 0.232 0.232
Qwen3.5-0.8B 0.230 0.230

img.embodied

model mean eb_alfred_visual
OneJev-27B 0.594 0.594
OneJev-27B-GGUF-Q8_0 0.592 0.592
Qwen3.8-27B-FP8 0.590 0.590
OneJev-9B 0.532 0.532
Qwen3.5-35B-A3B-GPTQ-Int4 0.504 0.504
OneJev-4B 0.592 0.592
Qwen3.5-9B 0.522 0.522
diffusiongemma-26B-A4B-it 0.552 0.552
Qwen3-Omni-30B-A3B-Thinking 0.494 0.494
Qwen3.5-4B 0.552 0.552
Qwen2.5-Omni-7B 0.508 0.508
Qwen2.5-Omni-7B-GPTQ-Int4 0.522 0.522
Qwen2.5-Omni-7B-AWQ 0.510 0.510
Qwen3-Omni-30B-A3B-Instruct 0.514 0.514
OneJev-0.8B 0.544 0.544
Qwen2.5-Omni-3B 0.528 0.528
Qwen3.5-0.8B 0.464 0.464

img.game

model mean atari_head board_twin glitch
OneJev-27B 0.423 0.158 0.469 0.642
OneJev-27B-GGUF-Q8_0 0.422 0.156 0.457 0.653
Qwen3.8-27B-FP8 0.412 0.150 0.413 0.674
OneJev-9B 0.414 0.182 0.462 0.598
Qwen3.5-35B-A3B-GPTQ-Int4 0.435 0.200 0.412 0.692
OneJev-4B 0.384 0.170 0.417 0.566
Qwen3.5-9B 0.402 0.182 0.361 0.664
diffusiongemma-26B-A4B-it 0.384 0.168 0.439 0.545
Qwen3-Omni-30B-A3B-Thinking 0.414 0.162 0.494 0.585
Qwen3.5-4B 0.395 0.174 0.349 0.662
Qwen2.5-Omni-7B 0.417 0.182 0.456 0.613
Qwen2.5-Omni-7B-GPTQ-Int4 0.420 0.184 0.440 0.636
Qwen2.5-Omni-7B-AWQ 0.413 0.158 0.450 0.630
Qwen3-Omni-30B-A3B-Instruct 0.353 0.188 0.379 0.493
OneJev-0.8B 0.374 0.196 0.411 0.515
Qwen2.5-Omni-3B 0.370 0.154 0.362 0.594
Qwen3.5-0.8B 0.293 0.154 0.304 0.422

img.judge

model mean agent_success hallucination preference t2i_alignment
OneJev-27B 0.809 0.836 0.814 0.714 0.873
OneJev-27B-GGUF-Q8_0 0.811 0.836 0.818 0.714 0.876
Qwen3.8-27B-FP8 0.813 0.778 0.778 0.860 0.835
OneJev-9B 0.786 0.844 0.782 0.644 0.876
Qwen3.5-35B-A3B-GPTQ-Int4 0.808 0.782 0.778 0.812 0.860
OneJev-4B 0.762 0.796 0.752 0.660 0.840
Qwen3.5-9B 0.739 0.626 0.749 0.758 0.822
diffusiongemma-26B-A4B-it 0.783 0.724 0.780 0.780 0.848
Qwen3-Omni-30B-A3B-Thinking 0.730 0.618 0.727 0.730 0.845
Qwen3.5-4B 0.708 0.684 0.711 0.652 0.783
Qwen2.5-Omni-7B 0.666 0.736 0.695 0.396 0.835
Qwen2.5-Omni-7B-GPTQ-Int4 0.688 0.738 0.695 0.484 0.835
Qwen2.5-Omni-7B-AWQ 0.661 0.746 0.681 0.402 0.817
Qwen3-Omni-30B-A3B-Instruct 0.693 0.576 0.661 0.750 0.783
OneJev-0.8B 0.689 0.762 0.635 0.606 0.752
Qwen2.5-Omni-3B 0.566 0.552 0.605 0.358 0.749
Qwen3.5-0.8B 0.509 0.508 0.621 0.260 0.646

img.know

model mean clock_calendar math_visual perception subject_mcq visual_rules
OneJev-27B 0.641 0.487 0.753 0.751 0.502 0.713
OneJev-27B-GGUF-Q8_0 0.643 0.484 0.758 0.750 0.506 0.720
Qwen3.8-27B-FP8 0.565 0.250 0.653 0.713 0.478 0.730
OneJev-9B 0.587 0.467 0.668 0.719 0.438 0.642
Qwen3.5-35B-A3B-GPTQ-Int4 0.560 0.256 0.652 0.733 0.488 0.670
OneJev-4B 0.564 0.408 0.662 0.694 0.382 0.675
Qwen3.5-9B 0.469 0.220 0.592 0.617 0.396 0.520
diffusiongemma-26B-A4B-it 0.503 0.204 0.637 0.606 0.426 0.640
Qwen3-Omni-30B-A3B-Thinking 0.462 0.133 0.569 0.666 0.386 0.557
Qwen3.5-4B 0.458 0.190 0.554 0.608 0.376 0.565
Qwen2.5-Omni-7B 0.451 0.113 0.613 0.566 0.344 0.618
Qwen2.5-Omni-7B-GPTQ-Int4 0.432 0.121 0.591 0.530 0.314 0.605
Qwen2.5-Omni-7B-AWQ 0.437 0.099 0.607 0.546 0.324 0.610
Qwen3-Omni-30B-A3B-Instruct 0.465 0.161 0.578 0.672 0.406 0.507
OneJev-0.8B 0.413 0.174 0.555 0.570 0.244 0.520
Qwen2.5-Omni-3B 0.395 0.067 0.561 0.511 0.254 0.585
Qwen3.5-0.8B 0.322 0.038 0.493 0.323 0.200 0.555

img.pro

model mean damage defect medical phishing_page
OneJev-27B 0.716 0.568 0.697 0.733 0.866
OneJev-27B-GGUF-Q8_0 0.715 0.567 0.701 0.731 0.862
Qwen3.8-27B-FP8 0.666 0.542 0.596 0.714 0.812
OneJev-9B 0.676 0.548 0.688 0.709 0.757
Qwen3.5-35B-A3B-GPTQ-Int4 0.692 0.565 0.689 0.718 0.796
OneJev-4B 0.657 0.563 0.671 0.687 0.705
Qwen3.5-9B 0.629 0.526 0.667 0.732 0.590
diffusiongemma-26B-A4B-it 0.670 0.560 0.656 0.693 0.771
Qwen3-Omni-30B-A3B-Thinking 0.590 0.487 0.660 0.668 0.546
Qwen3.5-4B 0.580 0.536 0.565 0.659 0.562
Qwen2.5-Omni-7B 0.501 0.404 0.450 0.609 0.541
Qwen2.5-Omni-7B-GPTQ-Int4 0.502 0.409 0.402 0.601 0.596
Qwen2.5-Omni-7B-AWQ 0.504 0.413 0.432 0.599 0.572
Qwen3-Omni-30B-A3B-Instruct 0.529 0.521 0.499 0.631 0.466
OneJev-0.8B 0.534 0.467 0.545 0.606 0.519
Qwen2.5-Omni-3B 0.480 0.411 0.446 0.565 0.498
Qwen3.5-0.8B 0.331 0.264 0.370 0.479 0.213

img.ret

model mean compositional same_item
OneJev-27B 0.915 0.920 0.910
OneJev-27B-GGUF-Q8_0 0.914 0.918 0.910
Qwen3.8-27B-FP8 0.873 0.908 0.838
OneJev-9B 0.867 0.910 0.824
Qwen3.5-35B-A3B-GPTQ-Int4 0.873 0.884 0.862
OneJev-4B 0.847 0.892 0.802
Qwen3.5-9B 0.888 0.884 0.892
diffusiongemma-26B-A4B-it 0.854 0.850 0.858
Qwen3-Omni-30B-A3B-Thinking 0.859 0.866 0.852
Qwen3.5-4B 0.870 0.860 0.880
Qwen2.5-Omni-7B 0.833 0.822 0.844
Qwen2.5-Omni-7B-GPTQ-Int4 0.842 0.822 0.862
Qwen2.5-Omni-7B-AWQ 0.804 0.774 0.834
Qwen3-Omni-30B-A3B-Instruct 0.855 0.842 0.868
OneJev-0.8B 0.804 0.784 0.824
Qwen2.5-Omni-3B 0.802 0.788 0.816
Qwen3.5-0.8B 0.650 0.578 0.722

img.route

model mean abstain document_routing model_correctness quality_gate
OneJev-27B 0.600 0.414 0.700 0.516 0.769
OneJev-27B-GGUF-Q8_0 0.600 0.414 0.704 0.512 0.768
Qwen3.8-27B-FP8 0.560 0.454 0.668 0.410 0.709
OneJev-9B 0.578 0.346 0.690 0.570 0.707
Qwen3.5-35B-A3B-GPTQ-Int4 0.586 0.406 0.714 0.518 0.707
OneJev-4B 0.590 0.342 0.672 0.586 0.761
Qwen3.5-9B 0.570 0.386 0.688 0.532 0.672
diffusiongemma-26B-A4B-it 0.535 0.350 0.584 0.514 0.691
Qwen3-Omni-30B-A3B-Thinking 0.580 0.380 0.652 0.542 0.746
Qwen3.5-4B 0.494 0.352 0.682 0.366 0.575
Qwen2.5-Omni-7B 0.481 0.200 0.580 0.386 0.759
Qwen2.5-Omni-7B-GPTQ-Int4 0.507 0.198 0.570 0.504 0.756
Qwen2.5-Omni-7B-AWQ 0.492 0.214 0.580 0.428 0.744
Qwen3-Omni-30B-A3B-Instruct 0.468 0.428 0.660 0.310 0.474
OneJev-0.8B 0.493 0.272 0.608 0.434 0.657
Qwen2.5-Omni-3B 0.466 0.238 0.566 0.394 0.666
Qwen3.5-0.8B 0.387 0.220 0.410 0.434 0.482

img.safety

model mean hateful_meme injection policy privacy synthetic
OneJev-27B 0.751 0.754 0.712 0.630 0.921 0.740
OneJev-27B-GGUF-Q8_0 0.748 0.742 0.708 0.624 0.921 0.744
Qwen3.8-27B-FP8 0.717 0.680 0.712 0.656 0.921 0.614
OneJev-9B 0.730 0.714 0.746 0.630 0.922 0.636
Qwen3.5-35B-A3B-GPTQ-Int4 0.741 0.718 0.681 0.598 0.908 0.800
OneJev-4B 0.712 0.710 0.637 0.603 0.912 0.698
Qwen3.5-9B 0.711 0.670 0.729 0.650 0.823 0.682
diffusiongemma-26B-A4B-it 0.708 (0.78) 0.712 0.698 0.600 0.820
Qwen3-Omni-30B-A3B-Thinking 0.661 0.642 0.627 0.591 0.914 0.530
Qwen3.5-4B 0.694 0.648 0.627 0.628 0.805 0.762
Qwen2.5-Omni-7B 0.698 0.646 0.668 0.635 0.929 0.612
Qwen2.5-Omni-7B-GPTQ-Int4 0.675 0.620 0.569 0.625 0.926 0.634
Qwen2.5-Omni-7B-AWQ 0.659 0.634 0.617 0.610 0.921 0.514
Qwen3-Omni-30B-A3B-Instruct 0.582 0.574 0.478 0.651 0.708 0.500
OneJev-0.8B 0.608 0.528 0.556 0.479 0.886 0.590
Qwen2.5-Omni-3B 0.638 0.608 0.607 0.592 0.926 0.458
Qwen3.5-0.8B 0.525 0.498 0.441 0.337 0.849 0.502

img.social

model mean art_emotion humor
OneJev-27B 0.293 0.406 0.180
OneJev-27B-GGUF-Q8_0 0.301 0.412 0.190
Qwen3.8-27B-FP8 0.374 0.352 0.396
OneJev-9B 0.293 0.374 0.212
Qwen3.5-35B-A3B-GPTQ-Int4 0.259 0.318 0.200
OneJev-4B 0.263 0.372 0.154
Qwen3.5-9B 0.235 0.334 0.136
diffusiongemma-26B-A4B-it 0.309 0.346 0.272
Qwen3-Omni-30B-A3B-Thinking 0.210 0.340 0.080
Qwen3.5-4B 0.419 0.362 0.476
Qwen2.5-Omni-7B 0.343 0.344 0.342
Qwen2.5-Omni-7B-GPTQ-Int4 0.329 0.326 0.332
Qwen2.5-Omni-7B-AWQ 0.361 0.354 0.368
Qwen3-Omni-30B-A3B-Instruct 0.315 0.340 0.290
OneJev-0.8B 0.339 0.310 0.368
Qwen2.5-Omni-3B 0.342 0.332 0.352
Qwen3.5-0.8B 0.136 0.226 0.046

img.std

model mean aokvqa gqa mmmu naturalbench okvqa scienceqa vizwiz_qa
OneJev-27B 0.859 0.890 0.813 0.690 0.843 0.948 0.968 0.862
OneJev-27B-GGUF-Q8_0
Qwen3.8-27B-FP8
OneJev-9B
Qwen3.5-35B-A3B-GPTQ-Int4
OneJev-4B
Qwen3.5-9B
diffusiongemma-26B-A4B-it
Qwen3-Omni-30B-A3B-Thinking
Qwen3.5-4B
Qwen2.5-Omni-7B
Qwen2.5-Omni-7B-GPTQ-Int4
Qwen2.5-Omni-7B-AWQ
Qwen3-Omni-30B-A3B-Instruct
OneJev-0.8B
Qwen2.5-Omni-3B
Qwen3.5-0.8B

img.step

model mean grounding gui_action
OneJev-27B 0.732 0.766 0.699
OneJev-27B-GGUF-Q8_0 0.730 0.766 0.695
Qwen3.8-27B-FP8 0.667 0.738 0.596
OneJev-9B 0.661 0.672 0.651
Qwen3.5-35B-A3B-GPTQ-Int4 0.605 0.718 0.492
OneJev-4B 0.649 0.660 0.639
Qwen3.5-9B 0.589 0.702 0.476
diffusiongemma-26B-A4B-it 0.549 0.650 0.448
Qwen3-Omni-30B-A3B-Thinking 0.502 0.586 0.418
Qwen3.5-4B 0.504 0.640 0.367
Qwen2.5-Omni-7B 0.465 0.582 0.347
Qwen2.5-Omni-7B-GPTQ-Int4 0.433 0.560 0.305
Qwen2.5-Omni-7B-AWQ 0.441 0.556 0.325
Qwen3-Omni-30B-A3B-Instruct 0.506 0.600 0.412
OneJev-0.8B 0.471 0.454 0.488
Qwen2.5-Omni-3B 0.364 0.484 0.245
Qwen3.5-0.8B 0.239 0.300 0.179

img.struct

model mean chart_table document_qa form_fields
OneJev-27B 0.889 0.841 0.843 0.982
OneJev-27B-GGUF-Q8_0 0.887 0.832 0.849 0.980
Qwen3.8-27B-FP8 0.861 0.774 0.842 0.968
OneJev-9B 0.839 0.824 0.806 0.888
Qwen3.5-35B-A3B-GPTQ-Int4 0.855 0.767 0.828 0.970
OneJev-4B 0.771 0.777 0.757 0.778
Qwen3.5-9B 0.815 0.747 0.777 0.922
diffusiongemma-26B-A4B-it 0.796 0.846 0.709 0.834
Qwen3-Omni-30B-A3B-Thinking 0.732 0.678 0.717 0.802
Qwen3.5-4B 0.721 0.668 0.680 0.816
Qwen2.5-Omni-7B 0.680 0.675 0.687 0.678
Qwen2.5-Omni-7B-GPTQ-Int4 0.644 0.648 0.663 0.622
Qwen2.5-Omni-7B-AWQ 0.667 0.675 0.667 0.660
Qwen3-Omni-30B-A3B-Instruct 0.762 0.685 0.705 0.896
OneJev-0.8B 0.500 0.595 0.556 0.350
Qwen2.5-Omni-3B 0.535 0.621 0.592 0.392
Qwen3.5-0.8B 0.406 0.526 0.427 0.266
  • Standard benchmarks (img.std, smoke run 2026-10-10, OneJev-27B only so far): A-OKVQA 0.890, OK-VQA 0.948, VizWiz-VQA 0.862, GQA 0.813, NaturalBench 0.843 (group accuracy 0.502), ScienceQA 0.968, MMMU 0.690. OK-VQA, VizWiz and GQA are open-answer benchmarks asked as four-option choices here, so they are not comparable with published open-answer accuracy. These leaves are not in the overall score.
  • OneJev beats its Qwen3.5 base at every size, and OneJev-27B beats Qwen3.8-27B-FP8 by 0.03.
  • The OneJev models are the best calibrated (ECE 0.077-0.099 against 0.11-0.30 for the others).
  • DiffusionGemma was not run on img.safety/privacy (coverage 0.978); its overall score is the mean of the other 39 leaves.
  • The scores come from the image team's pipeline runs, imported without new model calls. For img.know/subject_mcq, img.route/abstain and img.route/model_correctness those runs placed images inline in the question text.

Layout 

jevarena_mm/
  schema.py  registry.py  media.py  metrics.py  runner.py  cli.py     item format, discovery, scoring, CLI
  tasks/img/<domain>/<leaf>.py                                       one TaskAdapter per source
  models/qev.py                                                      adapters for models behind a qev endpoint
  models/servers/                                                    the servers those adapters call
data/img/<domain>/<leaf>/<source>.jsonl                              items: text + image references
data/img/<domain>/manifest.json                                      task cards, counts, file hashes, export provenance
scripts/fetch/<media_root>.py                                        images for one source
scripts/import_runs/                                                 existing predictions -> run cache (no model calls)
sweeps/*.toml                                                        run plans
results/summary/<release>/                                           committed result tables (results/runs/ is not)
docs/tasks_img_<domain>.md  docs/models.md  docs/input_examples.md  docs/failure_analysis.md  docs/open_decisions.md

Usage 

pip install -e '.[image]'                     # + '.[qev]' where models run
export JEVARENA_MM_MEDIA=/path/to/media      # images never go in Git
jevarena-mm tasks --leaf img.                 # registered tasks
jevarena-mm fetch --leaf img.                 # hosted folders from the HF dataset, the rest by scripts/fetch/
jevarena-mm verify --leaf img.                # schema, ids, manifest hashes, image sha256
jevarena-mm cards --leaf img.                 # docs/tasks_img_*.md
JEVARENA_MM_QEV_URL=http://localhost:8000 jevarena-mm run --model 'onejev@model=OneJev-27B,img=auto,debias=1' --leaf img.
jevarena-mm score --model 'onejev@model=OneJev-27B,img=auto,debias=1' --leaf img.
jevarena-mm compare                           # results/summary/<release>/README.md

build regenerates data/ from the task adapters. The image tasks read the image team's frozen export (JEVARENA_MM_IMG_EXPORT); the export comes from the team's pipeline, which stays internal.

Item format (mm-0.1) 

{"schema_version": "mm-0.1", "id": "img.basic.custom_categories.cub:e1b5054f69ff",
 "leaf": "img.basic/custom_categories", "task": "img.basic/custom_categories/cub",
 "request": {"state": {"task": "Which value of has breast pattern does this bird have?",
                       "image": {"kind": "image", "uri": "jevarena_img/img.basic_custom_categories/e6222a0b….jpg",
                                 "sha256": "e6222a0b…", "info": {"width": 500, "height": 333}}},
             "questions": {"attribute_value": {"type": "choice", "instructions": "…",
                                               "criteria": {"spotted": "spotted", "striped": "striped", "…": "…"}},
                           "has_attribute": {"type": "boolean", "instructions": "This bird has bill length: about the same as head"}}},
 "gold": {"attribute_value": {"label": "multi-colored", "dist": null}, "has_attribute": {"label": "true", "dist": null}},
 "meta": {"source": "CUB-200-2011, test split, attribute labels (Wah et al., 2011)", "license": "CC BY", "tier": "B",
          "group": "cub:3871", "text_twin": null, "media": [{"origin": "source", "source_path": "CUB_200_2011/images/…"}]}}
  • request is everything the model sees; gold and meta are never sent.
  • A state value is text, an image reference, or an ordered list of references; key order is presentation order.
  • Questions: choice (2–255 options), boolean, score (2–10 levels, criteria is the list of level texts, gold is the level index). Multi-label decisions are several booleans (privacy: 64 per item).
  • gold.dist: the annotators' distribution, where the source has one: {key: p} for choice, one p per level for score.
  • gold.accept (see docs/open_decisions.md): further option keys that count as correct.
  • The id hashes the request, the gold and every image sha256.

Metrics

jevarena_mm/metrics.py: accuracy (avg@k, maj@k), Brier (multi-class, sum over options of (p - t)^2 with t the annotators' distribution where the item has one, else one-hot; also reported one-hot everywhere), ECE (10 equal-mass bins), latency. Failures score 0; items not run are left out and every score is shown with its coverage. An item's questions are averaged first, a leaf is the mean of its items, and a domain or the overall score is the mean of its leaves.

Releases

Tags vMAJOR.MINOR.PATCH; item files are published to a Hugging Face dataset revision pinned by the tag's manifests. Only released tasks enter a version's tables. Never committed: images, model weights, secrets (.env), raw runs.

Removal

To have a file removed, open a discussion on this page naming the file; it is taken down.
Downloads last month
6
Total file size:
9.4 GB