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)

All 17 chat models were re-run on the original image files (2026-10-12). Responses whose model input stayed byte for byte the same were carried over; every other item was run again. Overall scores moved by at most 0.001 against the runs on the pipeline-built copies; the ranking is unchanged.

Seven System-One releases were added on 2026-10-14 (clef, clef-flash, clef-omni, Xor-26B-A4B v1.1, Jev-Omni, Jev-Omni-general-head-v3, raya-decision-v1). Each is served in its own format from its model card or bundled code, not through the letter-logit readout the chat models share; see "System-One releases" in docs/models.md for the formats and the deviations our items force. Clef and clef-flash fail the 14 mind2web items whose screenshot plus 255-option schema exceed Clef's default 16,384-token limit (scored as wrong).

Six more image-capable models were added on 2026-10-16: d3, Gemma 4 26B-A4B and 12B, Qwen3-VL-8B (run instead of 32B), Qwen3.5-2B and Phi-4-multimodal. d3 runs in its own sharded runtime; after a CUDA out-of-memory error it retries the batch one question at a time. 7 of its items still fail: 6 long mind2web items that run out of memory on two A100s (to be retried on three), and one A-OKVQA item whose criteria carry an empty key (rejected by d3's API, scored as wrong). Phi-4-multimodal, MiniCPM-o and Nemotron-3-Nano-Omni ship remote code written for transformers 4.x; the shims in jevarena_mm/models/servers/readout_remote.py adapt it to 5.x. The AWQ-INT4 DiffusionGemma was dropped: its packed experts load only in vLLM. MiniCPM-o 4.5 / 2.6 and Nemotron-3-Nano-Omni are still running.

All 30 model specs 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 d3 0.697 0.697 0.377 0.409 0.095 1.000
2 OneJev-27B-GGUF-Q8_0 0.678 0.678 0.388 0.417 0.079 1.000
3 OneJev-27B 0.678 0.678 0.388 0.417 0.077 1.000
4 clef 0.673 0.673 0.416 0.445 0.115 1.000
5 Qwen3.8-27B-FP8 0.648 0.648 0.439 0.467 0.112 1.000
6 Xor-26B-A4B-v1.1 0.647 0.647 0.439 0.470 0.134 1.000
7 gemma-4-26B-A4B-it 0.645 0.645 0.629 0.656 0.318 1.000
8 Qwen3.5-35B-A3B-GPTQ-Int4 0.645 0.645 0.461 0.490 0.148 1.000
9 OneJev-9B 0.645 0.645 0.426 0.455 0.084 1.000
10 clef-flash 0.628 0.628 0.465 0.494 0.113 1.000
11 OneJev-4B 0.624 0.624 0.444 0.473 0.082 1.000
12 clef-omni 0.620 0.620 0.488 0.516 0.139 1.000
13 gemma-4-12B-it 0.616 0.616 0.647 0.675 0.316 1.000
14 Qwen3.5-9B 0.607 0.607 0.492 0.522 0.150 1.000
15 diffusiongemma-26B-A4B-it 0.602 0.602 0.649 0.678 0.298 0.978
16 Qwen3-VL-8B-Instruct 0.594 0.594 0.652 0.683 0.298 1.000
17 Qwen3-Omni-30B-A3B-Thinking 0.575 0.575 0.666 0.696 0.289 1.000
18 Qwen3.5-4B 0.573 0.573 0.521 0.551 0.136 1.000
19 Jev-Omni 0.570 0.570 0.546 0.576 0.166 1.000
20 Qwen2.5-Omni-7B 0.554 0.554 0.568 0.594 0.170 1.000
21 Jev-Omni-general-head-v3 0.551 0.551 0.543 0.572 0.148 1.000
22 Qwen2.5-Omni-7B-GPTQ-Int4 0.547 0.547 0.560 0.586 0.153 1.000
23 Qwen2.5-Omni-7B-AWQ 0.545 0.545 0.570 0.596 0.168 1.000
24 Qwen3-Omni-30B-A3B-Instruct 0.543 0.543 0.672 0.701 0.278 1.000
25 OneJev-0.8B 0.518 0.518 0.548 0.578 0.101 1.000
26 Qwen2.5-Omni-3B 0.496 0.496 0.583 0.610 0.135 1.000
27 Qwen3.5-2B 0.466 0.466 0.647 0.678 0.199 1.000
28 raya-decision-v1 0.412 0.412 0.739 0.772 0.232 1.000
29 Qwen3.5-0.8B 0.388 0.388 0.695 0.728 0.210 1.000
30 Phi-4-multimodal-instruct 0.370 0.370 0.872 0.903 0.384 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
d3 0.701 0.639 0.610 0.630 0.408 0.814 0.746 0.710 0.921 0.639 0.719 0.277 0.872 0.762 0.925
OneJev-27B-GGUF-Q8_0 0.655 0.706 0.588 0.610 0.421 0.811 0.643 0.715 0.911 0.600 0.749 0.299 0.860 0.730 0.886
OneJev-27B 0.654 0.707 0.584 0.590 0.423 0.810 0.641 0.716 0.912 0.599 0.751 0.291 0.859 0.732 0.889
clef 0.684 0.693 0.522 0.568 0.434 0.779 0.674 0.642 0.915 0.625 0.740 0.325 0.851 0.692 0.892
Qwen3.8-27B-FP8 0.636 0.621 0.554 0.588 0.412 0.813 0.565 0.669 0.876 0.560 0.720 0.370 0.835 0.667 0.861
Xor-26B-A4B-v1.1 0.677 0.693 0.574 0.574 0.429 0.779 0.544 0.706 0.890 0.514 0.772 0.248 0.838 0.622 0.873
gemma-4-26B-A4B-it 0.639 0.694 0.476 0.624 0.386 0.775 0.531 0.716 0.878 0.596 0.743 0.319 0.827 0.611 0.868
Qwen3.5-35B-A3B-GPTQ-Int4 0.660 0.655 0.440 0.518 0.435 0.809 0.562 0.690 0.875 0.585 0.742 0.262 0.848 0.605 0.855
OneJev-9B 0.655 0.736 0.424 0.534 0.414 0.784 0.586 0.673 0.866 0.579 0.731 0.292 0.840 0.661 0.839
clef-flash 0.667 0.688 0.374 0.546 0.432 0.759 0.549 0.662 0.887 0.603 0.708 0.280 0.837 0.576 0.786
OneJev-4B 0.604 0.730 0.418 0.594 0.384 0.762 0.565 0.656 0.844 0.590 0.712 0.262 0.836 0.649 0.771
clef-omni 0.674 0.727 0.346 0.500 0.430 0.745 0.534 0.632 0.875 0.599 0.703 0.350 0.839 0.432 0.839
gemma-4-12B-it 0.568 0.684 0.474 0.596 0.412 0.748 0.505 0.688 0.873 0.570 0.708 0.272 0.789 0.586 0.792
Qwen3.5-9B 0.632 0.684 0.416 0.536 0.402 0.738 0.471 0.631 0.885 0.569 0.713 0.232 0.820 0.589 0.815
diffusiongemma-26B-A4B-it 0.550 0.544 0.470 0.570 0.384 0.783 0.499 0.668 0.850 0.535 0.709 (0.78) 0.317 0.805 0.549 0.796
Qwen3-VL-8B-Instruct 0.597 0.663 0.296 0.514 0.409 0.739 0.478 0.631 0.871 0.593 0.685 0.224 0.813 0.545 0.772
Qwen3-Omni-30B-A3B-Thinking 0.561 0.710 0.318 0.500 0.414 0.730 0.464 0.590 0.855 0.579 0.660 0.205 0.821 0.502 0.732
Qwen3.5-4B 0.538 0.532 0.402 0.538 0.395 0.713 0.460 0.582 0.868 0.494 0.692 0.419 0.810 0.504 0.721
Jev-Omni 0.509 0.692 0.482 0.494 0.387 0.751 0.457 0.592 0.807 0.501 0.702 0.237 0.773 0.434 0.767
Qwen2.5-Omni-7B 0.581 0.747 0.230 0.514 0.417 0.665 0.454 0.502 0.835 0.483 0.697 0.343 0.798 0.465 0.680
Jev-Omni-general-head-v3 0.499 0.654 0.514 0.518 0.377 0.714 0.435 0.574 0.814 0.460 0.679 0.267 0.781 0.405 0.731
Qwen2.5-Omni-7B-GPTQ-Int4 0.566 0.712 0.220 0.520 0.420 0.688 0.431 0.501 0.840 0.508 0.679 0.323 0.791 0.433 0.644
Qwen2.5-Omni-7B-AWQ 0.574 0.734 0.258 0.514 0.414 0.660 0.438 0.506 0.808 0.490 0.660 0.366 0.789 0.441 0.667
Qwen3-Omni-30B-A3B-Instruct 0.573 0.501 0.300 0.526 0.353 0.690 0.464 0.531 0.858 0.467 0.580 0.319 0.810 0.506 0.762
OneJev-0.8B 0.492 0.674 0.254 0.532 0.374 0.692 0.410 0.535 0.803 0.492 0.607 0.330 0.723 0.471 0.500
Qwen2.5-Omni-3B 0.471 0.639 0.232 0.556 0.370 0.574 0.399 0.480 0.809 0.467 0.638 0.337 0.764 0.364 0.535
Qwen3.5-2B 0.423 0.398 0.256 0.558 0.348 0.629 0.364 0.419 0.712 0.492 0.604 0.237 0.756 0.317 0.529
raya-decision-v1 0.326 0.671 0.148 0.504 0.285 0.651 0.335 0.402 0.606 0.400 0.580 0.138 0.658 0.164 0.369
Qwen3.5-0.8B 0.340 0.459 0.230 0.466 0.293 0.512 0.323 0.333 0.646 0.387 0.527 0.139 0.645 0.239 0.406
Phi-4-multimodal-instruct 0.297 0.269 0.228 0.500 0.302 0.532 0.343 0.329 0.618 0.247 0.438 0.386 0.558 0.212 0.416

By leaf

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

img.basic

model mean custom_categories fine_grained visual_entailment
d3 0.701 0.717 0.484 0.903
OneJev-27B-GGUF-Q8_0 0.655 0.698 0.408 0.859
OneJev-27B 0.654 0.698 0.404 0.859
clef 0.684 0.701 0.460 0.891
Qwen3.8-27B-FP8 0.636 0.697 0.358 0.852
Xor-26B-A4B-v1.1 0.677 0.697 0.498 0.836
gemma-4-26B-A4B-it 0.639 0.688 0.394 0.834
Qwen3.5-35B-A3B-GPTQ-Int4 0.660 0.658 0.470 0.852
OneJev-9B 0.655 0.685 0.460 0.819
clef-flash 0.667 0.673 0.466 0.862
OneJev-4B 0.604 0.664 0.350 0.799
clef-omni 0.674 0.691 0.462 0.869
gemma-4-12B-it 0.568 0.701 0.190 0.813
Qwen3.5-9B 0.632 0.659 0.412 0.825
diffusiongemma-26B-A4B-it 0.550 0.618 0.230 0.802
Qwen3-VL-8B-Instruct 0.597 0.681 0.286 0.825
Qwen3-Omni-30B-A3B-Thinking 0.561 0.568 0.364 0.750
Qwen3.5-4B 0.538 0.625 0.214 0.775
Jev-Omni 0.509 0.663 0.054 0.811
Qwen2.5-Omni-7B 0.581 0.581 0.380 0.782
Jev-Omni-general-head-v3 0.499 0.646 0.054 0.796
Qwen2.5-Omni-7B-GPTQ-Int4 0.566 0.585 0.352 0.760
Qwen2.5-Omni-7B-AWQ 0.574 0.581 0.358 0.782
Qwen3-Omni-30B-A3B-Instruct 0.573 0.601 0.408 0.709
OneJev-0.8B 0.492 0.628 0.182 0.665
Qwen2.5-Omni-3B 0.471 0.499 0.276 0.639
Qwen3.5-2B 0.423 0.511 0.078 0.681
raya-decision-v1 0.326 0.444 0.072 0.463
Qwen3.5-0.8B 0.340 0.520 0.028 0.472
Phi-4-multimodal-instruct 0.297 0.403 0.030 0.457

img.batch

model mean curation
d3 0.639 0.639
OneJev-27B-GGUF-Q8_0 0.706 0.706
OneJev-27B 0.707 0.707
clef 0.693 0.693
Qwen3.8-27B-FP8 0.621 0.621
Xor-26B-A4B-v1.1 0.693 0.693
gemma-4-26B-A4B-it 0.694 0.694
Qwen3.5-35B-A3B-GPTQ-Int4 0.655 0.655
OneJev-9B 0.736 0.736
clef-flash 0.688 0.688
OneJev-4B 0.730 0.730
clef-omni 0.727 0.727
gemma-4-12B-it 0.684 0.684
Qwen3.5-9B 0.684 0.684
diffusiongemma-26B-A4B-it 0.544 0.544
Qwen3-VL-8B-Instruct 0.663 0.663
Qwen3-Omni-30B-A3B-Thinking 0.710 0.710
Qwen3.5-4B 0.532 0.532
Jev-Omni 0.692 0.692
Qwen2.5-Omni-7B 0.747 0.747
Jev-Omni-general-head-v3 0.654 0.654
Qwen2.5-Omni-7B-GPTQ-Int4 0.712 0.712
Qwen2.5-Omni-7B-AWQ 0.734 0.734
Qwen3-Omni-30B-A3B-Instruct 0.501 0.501
OneJev-0.8B 0.674 0.674
Qwen2.5-Omni-3B 0.639 0.639
Qwen3.5-2B 0.398 0.398
raya-decision-v1 0.671 0.671
Qwen3.5-0.8B 0.459 0.459
Phi-4-multimodal-instruct 0.269 0.269

img.code

model mean render_match
d3 0.610 0.610
OneJev-27B-GGUF-Q8_0 0.588 0.588
OneJev-27B 0.584 0.584
clef 0.522 0.522
Qwen3.8-27B-FP8 0.554 0.554
Xor-26B-A4B-v1.1 0.574 0.574
gemma-4-26B-A4B-it 0.476 0.476
Qwen3.5-35B-A3B-GPTQ-Int4 0.440 0.440
OneJev-9B 0.424 0.424
clef-flash 0.374 0.374
OneJev-4B 0.418 0.418
clef-omni 0.346 0.346
gemma-4-12B-it 0.474 0.474
Qwen3.5-9B 0.416 0.416
diffusiongemma-26B-A4B-it 0.470 0.470
Qwen3-VL-8B-Instruct 0.296 0.296
Qwen3-Omni-30B-A3B-Thinking 0.318 0.318
Qwen3.5-4B 0.402 0.402
Jev-Omni 0.482 0.482
Qwen2.5-Omni-7B 0.230 0.230
Jev-Omni-general-head-v3 0.514 0.514
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
Qwen3.5-2B 0.256 0.256
raya-decision-v1 0.148 0.148
Qwen3.5-0.8B 0.230 0.230
Phi-4-multimodal-instruct 0.228 0.228

img.embodied

model mean eb_alfred_visual
d3 0.630 0.630
OneJev-27B-GGUF-Q8_0 0.610 0.610
OneJev-27B 0.590 0.590
clef 0.568 0.568
Qwen3.8-27B-FP8 0.588 0.588
Xor-26B-A4B-v1.1 0.574 0.574
gemma-4-26B-A4B-it 0.624 0.624
Qwen3.5-35B-A3B-GPTQ-Int4 0.518 0.518
OneJev-9B 0.534 0.534
clef-flash 0.546 0.546
OneJev-4B 0.594 0.594
clef-omni 0.500 0.500
gemma-4-12B-it 0.596 0.596
Qwen3.5-9B 0.536 0.536
diffusiongemma-26B-A4B-it 0.570 0.570
Qwen3-VL-8B-Instruct 0.514 0.514
Qwen3-Omni-30B-A3B-Thinking 0.500 0.500
Qwen3.5-4B 0.538 0.538
Jev-Omni 0.494 0.494
Qwen2.5-Omni-7B 0.514 0.514
Jev-Omni-general-head-v3 0.518 0.518
Qwen2.5-Omni-7B-GPTQ-Int4 0.520 0.520
Qwen2.5-Omni-7B-AWQ 0.514 0.514
Qwen3-Omni-30B-A3B-Instruct 0.526 0.526
OneJev-0.8B 0.532 0.532
Qwen2.5-Omni-3B 0.556 0.556
Qwen3.5-2B 0.558 0.558
raya-decision-v1 0.504 0.504
Qwen3.5-0.8B 0.466 0.466
Phi-4-multimodal-instruct 0.500 0.500

img.game

model mean atari_head board_twin glitch
d3 0.408 0.180 0.468 0.575
OneJev-27B-GGUF-Q8_0 0.421 0.154 0.457 0.653
OneJev-27B 0.423 0.158 0.469 0.642
clef 0.434 0.144 0.539 0.618
Qwen3.8-27B-FP8 0.412 0.150 0.413 0.674
Xor-26B-A4B-v1.1 0.429 0.184 0.536 0.567
gemma-4-26B-A4B-it 0.386 0.170 0.433 0.555
Qwen3.5-35B-A3B-GPTQ-Int4 0.435 0.200 0.412 0.692
OneJev-9B 0.414 0.182 0.462 0.598
clef-flash 0.432 0.176 0.567 0.554
OneJev-4B 0.384 0.170 0.417 0.566
clef-omni 0.430 0.180 0.583 0.526
gemma-4-12B-it 0.412 0.186 0.398 0.652
Qwen3.5-9B 0.402 0.182 0.361 0.664
diffusiongemma-26B-A4B-it 0.384 0.168 0.439 0.545
Qwen3-VL-8B-Instruct 0.409 0.160 0.470 0.598
Qwen3-Omni-30B-A3B-Thinking 0.414 0.162 0.494 0.585
Qwen3.5-4B 0.395 0.174 0.349 0.662
Jev-Omni 0.387 0.172 0.334 0.654
Qwen2.5-Omni-7B 0.417 0.182 0.456 0.613
Jev-Omni-general-head-v3 0.377 0.178 0.331 0.623
Qwen2.5-Omni-7B-GPTQ-Int4 0.420 0.184 0.440 0.636
Qwen2.5-Omni-7B-AWQ 0.414 0.160 0.450 0.632
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
Qwen3.5-2B 0.348 0.142 0.362 0.539
raya-decision-v1 0.285 0.160 0.276 0.419
Qwen3.5-0.8B 0.293 0.154 0.304 0.422
Phi-4-multimodal-instruct 0.302 0.158 0.335 0.413

img.judge

model mean agent_success hallucination preference t2i_alignment
d3 0.814 0.744 0.840 0.788 0.886
OneJev-27B-GGUF-Q8_0 0.811 0.836 0.818 0.718 0.871
OneJev-27B 0.810 0.836 0.814 0.716 0.873
clef 0.779 0.750 0.800 0.680 0.886
Qwen3.8-27B-FP8 0.813 0.778 0.778 0.858 0.837
Xor-26B-A4B-v1.1 0.779 0.702 0.774 0.804 0.837
gemma-4-26B-A4B-it 0.775 0.754 0.780 0.724 0.842
Qwen3.5-35B-A3B-GPTQ-Int4 0.809 0.782 0.778 0.818 0.858
OneJev-9B 0.784 0.844 0.782 0.644 0.866
clef-flash 0.759 0.758 0.786 0.622 0.871
OneJev-4B 0.762 0.796 0.752 0.664 0.835
clef-omni 0.745 0.722 0.758 0.652 0.850
gemma-4-12B-it 0.748 0.758 0.766 0.606 0.863
Qwen3.5-9B 0.738 0.626 0.749 0.762 0.814
diffusiongemma-26B-A4B-it 0.783 0.724 0.780 0.774 0.853
Qwen3-VL-8B-Instruct 0.739 0.614 0.758 0.772 0.811
Qwen3-Omni-30B-A3B-Thinking 0.730 0.618 0.727 0.730 0.845
Qwen3.5-4B 0.713 0.684 0.711 0.672 0.786
Jev-Omni 0.751 0.730 0.754 0.692 0.827
Qwen2.5-Omni-7B 0.665 0.736 0.695 0.388 0.842
Jev-Omni-general-head-v3 0.714 0.638 0.737 0.684 0.796
Qwen2.5-Omni-7B-GPTQ-Int4 0.688 0.738 0.695 0.484 0.835
Qwen2.5-Omni-7B-AWQ 0.660 0.746 0.681 0.400 0.811
Qwen3-Omni-30B-A3B-Instruct 0.690 0.576 0.661 0.744 0.780
OneJev-0.8B 0.692 0.762 0.635 0.612 0.757
Qwen2.5-Omni-3B 0.574 0.552 0.605 0.392 0.747
Qwen3.5-2B 0.629 0.540 0.683 0.574 0.718
raya-decision-v1 0.651 0.548 0.587 0.772 0.695
Qwen3.5-0.8B 0.512 0.508 0.621 0.272 0.646
Phi-4-multimodal-instruct 0.532 0.510 0.435 0.498 0.685

img.know

model mean clock_calendar math_visual perception subject_mcq visual_rules
d3 0.746 0.797 0.851 0.787 0.596 0.698
OneJev-27B-GGUF-Q8_0 0.643 0.484 0.759 0.753 0.502 0.718
OneJev-27B 0.641 0.487 0.754 0.748 0.502 0.715
clef 0.674 0.681 0.755 0.776 0.502 0.657
Qwen3.8-27B-FP8 0.565 0.250 0.652 0.712 0.478 0.733
Xor-26B-A4B-v1.1 0.544 0.292 0.671 0.727 0.532 0.500
gemma-4-26B-A4B-it 0.531 0.210 0.665 0.632 0.478 0.672
Qwen3.5-35B-A3B-GPTQ-Int4 0.562 0.256 0.654 0.733 0.486 0.682
OneJev-9B 0.586 0.467 0.668 0.714 0.436 0.645
clef-flash 0.549 0.424 0.731 0.651 0.432 0.507
OneJev-4B 0.565 0.408 0.658 0.700 0.384 0.677
clef-omni 0.534 0.350 0.680 0.679 0.462 0.497
gemma-4-12B-it 0.505 0.228 0.639 0.620 0.412 0.627
Qwen3.5-9B 0.471 0.220 0.593 0.619 0.396 0.525
diffusiongemma-26B-A4B-it 0.499 0.204 0.635 0.610 0.432 0.615
Qwen3-VL-8B-Instruct 0.478 0.128 0.630 0.603 0.366 0.665
Qwen3-Omni-30B-A3B-Thinking 0.464 0.133 0.570 0.670 0.388 0.560
Qwen3.5-4B 0.460 0.190 0.554 0.613 0.378 0.568
Jev-Omni 0.457 0.182 0.629 0.580 0.390 0.502
Qwen2.5-Omni-7B 0.454 0.113 0.613 0.570 0.346 0.627
Jev-Omni-general-head-v3 0.435 0.168 0.601 0.531 0.376 0.500
Qwen2.5-Omni-7B-GPTQ-Int4 0.431 0.121 0.591 0.530 0.312 0.603
Qwen2.5-Omni-7B-AWQ 0.438 0.099 0.607 0.549 0.324 0.610
Qwen3-Omni-30B-A3B-Instruct 0.464 0.161 0.576 0.671 0.404 0.507
OneJev-0.8B 0.410 0.174 0.556 0.560 0.246 0.515
Qwen2.5-Omni-3B 0.399 0.067 0.560 0.516 0.254 0.600
Qwen3.5-2B 0.364 0.057 0.513 0.506 0.246 0.500
raya-decision-v1 0.335 0.063 0.513 0.442 0.156 0.502
Qwen3.5-0.8B 0.323 0.038 0.492 0.323 0.202 0.562
Phi-4-multimodal-instruct 0.343 0.063 0.444 0.458 0.248 0.502

img.pro

model mean damage defect medical phishing_page
d3 0.710 0.565 0.722 0.765 0.789
OneJev-27B-GGUF-Q8_0 0.715 0.567 0.698 0.734 0.862
OneJev-27B 0.716 0.565 0.701 0.732 0.866
clef 0.642 0.566 0.649 0.769 0.583
Qwen3.8-27B-FP8 0.669 0.547 0.600 0.716 0.812
Xor-26B-A4B-v1.1 0.706 0.563 0.701 0.741 0.818
gemma-4-26B-A4B-it 0.716 0.595 0.728 0.727 0.812
Qwen3.5-35B-A3B-GPTQ-Int4 0.690 0.567 0.689 0.707 0.796
OneJev-9B 0.673 0.549 0.683 0.705 0.757
clef-flash 0.662 0.556 0.650 0.736 0.705
OneJev-4B 0.656 0.564 0.670 0.684 0.705
clef-omni 0.632 0.592 0.627 0.726 0.583
gemma-4-12B-it 0.688 0.552 0.700 0.680 0.821
Qwen3.5-9B 0.631 0.530 0.674 0.732 0.590
diffusiongemma-26B-A4B-it 0.668 0.552 0.666 0.685 0.771
Qwen3-VL-8B-Instruct 0.631 0.500 0.662 0.626 0.736
Qwen3-Omni-30B-A3B-Thinking 0.590 0.488 0.658 0.668 0.546
Qwen3.5-4B 0.582 0.539 0.566 0.662 0.562
Jev-Omni 0.592 0.449 0.638 0.696 0.585
Qwen2.5-Omni-7B 0.502 0.410 0.449 0.609 0.541
Jev-Omni-general-head-v3 0.574 0.464 0.603 0.681 0.547
Qwen2.5-Omni-7B-GPTQ-Int4 0.501 0.414 0.394 0.599 0.596
Qwen2.5-Omni-7B-AWQ 0.506 0.416 0.436 0.599 0.572
Qwen3-Omni-30B-A3B-Instruct 0.531 0.517 0.508 0.634 0.466
OneJev-0.8B 0.535 0.469 0.547 0.606 0.519
Qwen2.5-Omni-3B 0.480 0.412 0.452 0.559 0.498
Qwen3.5-2B 0.419 0.402 0.397 0.575 0.301
raya-decision-v1 0.402 0.260 0.406 0.522 0.421
Qwen3.5-0.8B 0.333 0.268 0.370 0.480 0.213
Phi-4-multimodal-instruct 0.329 0.292 0.392 0.332 0.301

img.ret

model mean compositional same_item
d3 0.921 0.936 0.906
OneJev-27B-GGUF-Q8_0 0.911 0.918 0.904
OneJev-27B 0.912 0.920 0.904
clef 0.915 0.918 0.912
Qwen3.8-27B-FP8 0.876 0.908 0.844
Xor-26B-A4B-v1.1 0.890 0.926 0.854
gemma-4-26B-A4B-it 0.878 0.906 0.850
Qwen3.5-35B-A3B-GPTQ-Int4 0.875 0.884 0.866
OneJev-9B 0.866 0.910 0.822
clef-flash 0.887 0.906 0.868
OneJev-4B 0.844 0.892 0.796
clef-omni 0.875 0.904 0.846
gemma-4-12B-it 0.873 0.880 0.866
Qwen3.5-9B 0.885 0.884 0.886
diffusiongemma-26B-A4B-it 0.850 0.850 0.850
Qwen3-VL-8B-Instruct 0.871 0.872 0.870
Qwen3-Omni-30B-A3B-Thinking 0.855 0.866 0.844
Qwen3.5-4B 0.868 0.860 0.876
Jev-Omni 0.807 0.870 0.744
Qwen2.5-Omni-7B 0.835 0.822 0.848
Jev-Omni-general-head-v3 0.814 0.868 0.760
Qwen2.5-Omni-7B-GPTQ-Int4 0.840 0.822 0.858
Qwen2.5-Omni-7B-AWQ 0.808 0.774 0.842
Qwen3-Omni-30B-A3B-Instruct 0.858 0.842 0.874
OneJev-0.8B 0.803 0.784 0.822
Qwen2.5-Omni-3B 0.809 0.788 0.830
Qwen3.5-2B 0.712 0.750 0.674
raya-decision-v1 0.606 0.712 0.500
Qwen3.5-0.8B 0.646 0.578 0.714
Phi-4-multimodal-instruct 0.618 0.610 0.626

img.route

model mean abstain document_routing model_correctness quality_gate
d3 0.639 0.376 0.720 0.676 0.784
OneJev-27B-GGUF-Q8_0 0.600 0.414 0.704 0.512 0.771
OneJev-27B 0.599 0.414 0.700 0.516 0.767
clef 0.625 0.348 0.670 0.698 0.785
Qwen3.8-27B-FP8 0.560 0.454 0.668 0.410 0.709
Xor-26B-A4B-v1.1 0.514 0.308 0.722 0.314 0.714
gemma-4-26B-A4B-it 0.596 0.398 0.702 0.546 0.738
Qwen3.5-35B-A3B-GPTQ-Int4 0.585 0.406 0.714 0.518 0.704
OneJev-9B 0.579 0.346 0.690 0.570 0.708
clef-flash 0.603 0.300 0.680 0.682 0.749
OneJev-4B 0.590 0.342 0.672 0.586 0.761
clef-omni 0.599 0.346 0.670 0.614 0.767
gemma-4-12B-it 0.570 0.334 0.718 0.476 0.751
Qwen3.5-9B 0.569 0.386 0.688 0.532 0.672
diffusiongemma-26B-A4B-it 0.535 0.350 0.584 0.514 0.692
Qwen3-VL-8B-Instruct 0.593 0.408 0.680 0.586 0.697
Qwen3-Omni-30B-A3B-Thinking 0.579 0.380 0.652 0.542 0.743
Qwen3.5-4B 0.494 0.352 0.682 0.366 0.575
Jev-Omni 0.501 0.242 0.652 0.384 0.725
Qwen2.5-Omni-7B 0.483 0.200 0.580 0.386 0.765
Jev-Omni-general-head-v3 0.460 0.236 0.620 0.326 0.657
Qwen2.5-Omni-7B-GPTQ-Int4 0.508 0.198 0.570 0.504 0.759
Qwen2.5-Omni-7B-AWQ 0.490 0.214 0.580 0.428 0.739
Qwen3-Omni-30B-A3B-Instruct 0.467 0.428 0.660 0.310 0.469
OneJev-0.8B 0.492 0.272 0.608 0.434 0.655
Qwen2.5-Omni-3B 0.467 0.238 0.566 0.394 0.671
Qwen3.5-2B 0.492 0.206 0.538 0.634 0.592
raya-decision-v1 0.400 0.232 0.288 0.598 0.483
Qwen3.5-0.8B 0.387 0.220 0.410 0.434 0.483
Phi-4-multimodal-instruct 0.247 0.232 0.066 0.304 0.386

img.safety

model mean hateful_meme injection policy privacy synthetic
d3 0.719 0.732 0.688 0.641 0.940 0.596
OneJev-27B-GGUF-Q8_0 0.749 0.740 0.708 0.631 0.921 0.744
OneJev-27B 0.751 0.748 0.712 0.631 0.922 0.740
clef 0.740 0.642 0.746 0.636 0.953 0.726
Qwen3.8-27B-FP8 0.720 0.692 0.712 0.661 0.921 0.614
Xor-26B-A4B-v1.1 0.772 0.724 0.732 0.607 0.927 0.872
gemma-4-26B-A4B-it 0.743 0.702 0.698 0.613 0.925 0.776
Qwen3.5-35B-A3B-GPTQ-Int4 0.742 0.718 0.681 0.600 0.909 0.800
OneJev-9B 0.731 0.716 0.746 0.633 0.922 0.636
clef-flash 0.708 0.626 0.705 0.628 0.945 0.634
OneJev-4B 0.712 0.704 0.637 0.611 0.912 0.698
clef-omni 0.703 0.662 0.671 0.612 0.950 0.618
gemma-4-12B-it 0.708 0.702 0.654 0.605 0.911 0.670
Qwen3.5-9B 0.713 0.674 0.729 0.655 0.823 0.682
diffusiongemma-26B-A4B-it 0.709 (0.78) 0.700 0.698 0.617 0.820
Qwen3-VL-8B-Instruct 0.685 0.626 0.607 0.613 0.903 0.674
Qwen3-Omni-30B-A3B-Thinking 0.660 0.636 0.627 0.594 0.913 0.530
Qwen3.5-4B 0.692 0.640 0.627 0.626 0.805 0.762
Jev-Omni 0.702 0.648 0.725 0.593 0.901 0.644
Qwen2.5-Omni-7B 0.697 0.642 0.668 0.632 0.929 0.612
Jev-Omni-general-head-v3 0.679 0.652 0.719 0.581 0.880 0.562
Qwen2.5-Omni-7B-GPTQ-Int4 0.679 0.640 0.569 0.627 0.926 0.634
Qwen2.5-Omni-7B-AWQ 0.660 0.638 0.617 0.608 0.921 0.514
Qwen3-Omni-30B-A3B-Instruct 0.580 0.574 0.478 0.640 0.708 0.500
OneJev-0.8B 0.607 0.522 0.556 0.479 0.888 0.590
Qwen2.5-Omni-3B 0.638 0.618 0.607 0.582 0.926 0.458
Qwen3.5-2B 0.604 0.528 0.515 0.511 0.884 0.580
raya-decision-v1 0.580 0.560 0.498 0.382 0.927 0.534
Qwen3.5-0.8B 0.527 0.506 0.441 0.340 0.848 0.502
Phi-4-multimodal-instruct 0.438 0.518 0.502 0.475 0.197 0.500

img.social

model mean art_emotion humor
d3 0.277 0.312 0.242
OneJev-27B-GGUF-Q8_0 0.299 0.402 0.196
OneJev-27B 0.291 0.400 0.182
clef 0.325 0.326 0.324
Qwen3.8-27B-FP8 0.370 0.342 0.398
Xor-26B-A4B-v1.1 0.248 0.338 0.158
gemma-4-26B-A4B-it 0.319 0.360 0.278
Qwen3.5-35B-A3B-GPTQ-Int4 0.262 0.324 0.200
OneJev-9B 0.292 0.372 0.212
clef-flash 0.280 0.298 0.262
OneJev-4B 0.262 0.368 0.156
clef-omni 0.350 0.320 0.380
gemma-4-12B-it 0.272 0.312 0.232
Qwen3.5-9B 0.232 0.334 0.130
diffusiongemma-26B-A4B-it 0.317 0.360 0.274
Qwen3-VL-8B-Instruct 0.224 0.304 0.144
Qwen3-Omni-30B-A3B-Thinking 0.205 0.332 0.078
Qwen3.5-4B 0.419 0.356 0.482
Jev-Omni 0.237 0.300 0.174
Qwen2.5-Omni-7B 0.343 0.340 0.346
Jev-Omni-general-head-v3 0.267 0.298 0.236
Qwen2.5-Omni-7B-GPTQ-Int4 0.323 0.322 0.324
Qwen2.5-Omni-7B-AWQ 0.366 0.360 0.372
Qwen3-Omni-30B-A3B-Instruct 0.319 0.354 0.284
OneJev-0.8B 0.330 0.300 0.360
Qwen2.5-Omni-3B 0.337 0.322 0.352
Qwen3.5-2B 0.237 0.290 0.184
raya-decision-v1 0.138 0.228 0.048
Qwen3.5-0.8B 0.139 0.232 0.046
Phi-4-multimodal-instruct 0.386 0.166 0.606

img.std

model mean aokvqa gqa mmmu naturalbench okvqa scienceqa vizwiz_qa
d3 0.872 0.884 0.842 0.728 0.860 0.960 0.984 0.846
OneJev-27B-GGUF-Q8_0 0.860 0.896 0.816 0.688 0.842 0.948 0.966 0.862
OneJev-27B 0.859 0.890 0.813 0.690 0.843 0.948 0.968 0.862
clef 0.851 0.896 0.809 0.632 0.834 0.954 0.968 0.866
Qwen3.8-27B-FP8 0.835 0.876 0.803 0.626 0.831 0.934 0.948 0.830
Xor-26B-A4B-v1.1 0.838 0.886 0.785 0.664 0.830 0.942 0.912 0.844
gemma-4-26B-A4B-it 0.827 0.860 0.779 0.648 0.823 0.944 0.928 0.804
Qwen3.5-35B-A3B-GPTQ-Int4 0.848 0.900 0.785 0.662 0.842 0.938 0.966 0.840
OneJev-9B 0.840 0.880 0.803 0.606 0.834 0.942 0.954 0.860
clef-flash 0.837 0.888 0.768 0.620 0.834 0.950 0.974 0.824
OneJev-4B 0.836 0.854 0.805 0.596 0.825 0.930 0.956 0.884
clef-omni 0.839 0.886 0.796 0.630 0.808 0.940 0.974 0.840
gemma-4-12B-it 0.789 0.784 0.751 0.620 0.810 0.900 0.832 0.824
Qwen3.5-9B 0.820 0.858 0.764 0.600 0.826 0.936 0.932 0.828
diffusiongemma-26B-A4B-it 0.805 0.826 0.781 0.592 0.806 0.908 0.900 0.822
Qwen3-VL-8B-Instruct 0.813 0.866 0.759 0.579 0.817 0.924 0.910 0.834
Qwen3-Omni-30B-A3B-Thinking 0.821 0.874 0.755 0.606 0.821 0.932 0.932 0.828
Qwen3.5-4B 0.810 0.852 0.768 0.590 0.808 0.912 0.920 0.818
Jev-Omni 0.773 0.774 0.764 0.559 0.812 0.878 0.808 0.816
Qwen2.5-Omni-7B 0.798 0.858 0.768 0.541 0.782 0.938 0.906 0.796
Jev-Omni-general-head-v3 0.781 0.790 0.757 0.549 0.805 0.886 0.860 0.818
Qwen2.5-Omni-7B-GPTQ-Int4 0.791 0.848 0.770 0.517 0.777 0.934 0.884 0.808
Qwen2.5-Omni-7B-AWQ 0.789 0.836 0.757 0.533 0.794 0.930 0.892 0.784
Qwen3-Omni-30B-A3B-Instruct 0.810 0.870 0.738 0.590 0.811 0.916 0.940 0.804
OneJev-0.8B 0.723 0.758 0.727 0.404 0.774 0.862 0.758 0.780
Qwen2.5-Omni-3B 0.764 0.824 0.777 0.437 0.762 0.920 0.826 0.806
Qwen3.5-2B 0.756 0.796 0.753 0.479 0.778 0.894 0.810 0.784
raya-decision-v1 0.658 0.735 0.757 0.314 0.703 0.854 0.626 0.614
Qwen3.5-0.8B 0.645 0.723 0.651 0.370 0.719 0.828 0.656 0.568
Phi-4-multimodal-instruct 0.558 0.543 0.651 0.380 0.598 0.594 0.694 0.442

img.step

model mean grounding gui_action
d3 0.762 0.794 0.731
OneJev-27B-GGUF-Q8_0 0.730 0.766 0.695
OneJev-27B 0.732 0.766 0.699
clef 0.692 0.754 0.631
Qwen3.8-27B-FP8 0.667 0.738 0.596
Xor-26B-A4B-v1.1 0.622 0.672 0.572
gemma-4-26B-A4B-it 0.611 0.694 0.528
Qwen3.5-35B-A3B-GPTQ-Int4 0.605 0.718 0.492
OneJev-9B 0.661 0.672 0.651
clef-flash 0.576 0.618 0.534
OneJev-4B 0.649 0.660 0.639
clef-omni 0.432 0.482 0.382
gemma-4-12B-it 0.586 0.646 0.526
Qwen3.5-9B 0.589 0.702 0.476
diffusiongemma-26B-A4B-it 0.549 0.650 0.448
Qwen3-VL-8B-Instruct 0.545 0.586 0.504
Qwen3-Omni-30B-A3B-Thinking 0.502 0.586 0.418
Qwen3.5-4B 0.504 0.640 0.367
Jev-Omni 0.434 0.546 0.321
Qwen2.5-Omni-7B 0.465 0.582 0.347
Jev-Omni-general-head-v3 0.405 0.500 0.309
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
Qwen3.5-2B 0.317 0.392 0.243
raya-decision-v1 0.164 0.176 0.153
Qwen3.5-0.8B 0.239 0.300 0.179
Phi-4-multimodal-instruct 0.212 0.252 0.173

img.struct

model mean chart_table document_qa form_fields
d3 0.925 0.893 0.922 0.960
OneJev-27B-GGUF-Q8_0 0.886 0.832 0.847 0.980
OneJev-27B 0.889 0.841 0.843 0.982
clef 0.892 0.824 0.893 0.958
Qwen3.8-27B-FP8 0.861 0.774 0.842 0.968
Xor-26B-A4B-v1.1 0.873 0.821 0.815 0.984
gemma-4-26B-A4B-it 0.868 0.876 0.789 0.938
Qwen3.5-35B-A3B-GPTQ-Int4 0.855 0.767 0.828 0.970
OneJev-9B 0.839 0.824 0.806 0.888
clef-flash 0.786 0.804 0.869 0.686
OneJev-4B 0.771 0.777 0.757 0.778
clef-omni 0.839 0.791 0.837 0.890
gemma-4-12B-it 0.792 0.819 0.694 0.862
Qwen3.5-9B 0.815 0.747 0.777 0.922
diffusiongemma-26B-A4B-it 0.796 0.846 0.709 0.834
Qwen3-VL-8B-Instruct 0.772 0.750 0.729 0.836
Qwen3-Omni-30B-A3B-Thinking 0.732 0.678 0.717 0.802
Qwen3.5-4B 0.721 0.668 0.680 0.816
Jev-Omni 0.767 0.792 0.684 0.826
Qwen2.5-Omni-7B 0.680 0.675 0.687 0.678
Jev-Omni-general-head-v3 0.731 0.759 0.639 0.796
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
Qwen3.5-2B 0.529 0.591 0.535 0.460
raya-decision-v1 0.369 0.496 0.456 0.154
Qwen3.5-0.8B 0.406 0.526 0.427 0.266
Phi-4-multimodal-instruct 0.416 0.521 0.398 0.330
  • Standard benchmarks (img.std, all 17 models): OneJev-27B averages 0.859 over the seven (ScienceQA 0.968, OK-VQA 0.948, A-OKVQA 0.890, VizWiz-VQA 0.862, NaturalBench 0.843, GQA 0.813, MMMU 0.690), ahead of Qwen3.5-35B-A3B-GPTQ (0.848) and Qwen3.8-27B-FP8 (0.835). 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
63
Total file size:
9.4 GB