SuperGlance-QA
Paper: arXiv:2611.19138 · PDF · Project page: guobapei.github.io/SuperGlance · Code: GuobaPei/SuperGlance
The release terms of this repository are not decided yet. Each source keeps its own terms; see Sources and terms.
These are the multiple-choice questions that SuperGlance's stage 1 and stage 2 trained on. There is one JSON row per question, in the trainers' own format. The models trained on them are SuperGlance-9B-Stage1 (after stage 1), SuperGlance-9B-Stage2 (after stage 2) and SuperGlance-9B (the final model).
The images are not hosted here. Each row names its image in the source dataset. build/build_images.py writes each image from there and checks it against the SHA-256 of the file the trainer read.
Contents
| stage | questions in the training pool | questions here | sources here (of all) |
|---|---|---|---|
| 1 | 1,158,150 | 655,715 | 20 (24) |
| 2 | 1,240,040 | 737,605 | 22 (26) |
Questions used as published, with the source's own question and options, are not here: all of SAT, SIQA, PIQA and AI2D, and the yes/no questions of GQA and ChartQA. See Used as published, not here. counts.json gives, for every source and stage, the questions in the training pool, the questions here, and where the others are.
| path | content |
|---|---|
stage1/train/<source>.jsonl, stage2/train/<source>.jsonl |
the rows, one file per source |
counts.json |
per source and stage: questions in the training pool, questions here, and where the others are |
SCHEMA.md |
every field, and how to rebuild a trainer's data directory |
images/text/blank.png |
the blank image that every text-only question uses |
build/ |
build_images.py and its README: writes the other images from their source datasets |
Stage 2 contains the stage-1 questions without text_nota and openimages_vertical, plus 124,378 DriveLM-nuScenes questions. In 3,641 openimages_box questions the question text was reworded for stage 2.
The evaluation rows are not here. This leaves out the dev split we used to select checkpoints and our DriveLM test split.
Loading
import glob, json
rows = [json.loads(line) for f in sorted(glob.glob("stage1/train/*.jsonl")) for line in open(f)]
rows.sort(key=lambda r: r["row"]) # the trainer's order
SCHEMA.md describes the fields. build/README.md covers the images.
Sources and terms
Each source's terms are taken from its own page.
| file | stage 1 | stage 2 | questions from | terms |
|---|---|---|---|---|
tallyqa |
119,859 | 119,859 | TallyQA (via The Cauldron) | Apache-2.0 (manoja328/TallyQA_dataset) |
textvqa |
2,418 | 2,418 | TextVQA (via The Cauldron) | CC BY 4.0 (facebook/textvqa) |
openimages_box |
17,966 | 17,966 | built by us from OpenImages boxes | annotations CC BY 4.0, images CC BY 2.0 (Open Images V7) |
openimages_vertical |
2,488 | built by us from OpenImages boxes | as above | |
spatialsense |
10,339 | 10,339 | SpatialSense | BSD-2-Clause (princeton-vl/SpatialSense) |
drivelm_object |
86,373 | DriveLM-nuScenes, train split | CC BY-NC-SA 4.0 (OpenDriveLab/DriveLM); images: nuScenes terms of use | |
drivelm_ego |
15,680 | as above | as above | |
drivelm_relation |
11,352 | as above | as above | |
drivelm_box |
10,973 | as above | as above | |
text_csqa |
9,619 | 9,619 | CommonsenseQA | MIT (tau/commonsense_qa) |
text_cosmosqa |
25,223 | 25,223 | Cosmos QA | CC BY 4.0 (allenai/cosmos_qa) |
text_medmcqa |
40,000 | 40,000 | MedMCQA | MIT (medmcqa/medmcqa) |
text_medqa |
9,854 | 9,854 | MedQA (USMLE) | MIT (jind11/MedQA) |
text_qasc |
8,132 | 8,132 | QASC | Apache-2.0 (allenai/qasc); the HF card says CC BY 4.0 |
text_obqa |
4,945 | 4,945 | OpenBookQA | Apache-2.0 (allenai/OpenBookQA) |
text_aqua |
40,000 | 40,000 | AQuA-RAT | Apache-2.0 (deepmind/aqua_rat) |
text_wanli |
40,000 | 40,000 | WANLI | CC BY 4.0 (alisawuffles/WANLI) |
text_art |
40,000 | 40,000 | αNLI (ART) | Apache-2.0 (allenai/abductive-commonsense-reasoning) |
text_strategyqa |
2,290 | 2,290 | StrategyQA | MIT (eladsegal/strategyqa) |
text_ruletaker |
40,000 | 40,000 | RuleTaker | Apache-2.0 (allenai/ruletaker) |
GQA |
141,004 of 447,983 | 141,004 of 447,983 | GQA's open questions with our distractors, and our relation flips of its yes questions | the official GQA pages name no licence |
chartqa |
21,579 of 22,136 | 21,579 of 22,136 | ChartQA (via The Cauldron) with our numeric distractors | GPL-3.0 (vis-nlp/ChartQA) |
text_yesno |
39,999 | 39,999 | yes/no questions built by us from the text sources | each row follows its parent source; 6,148 rows come from SIQA (no licence found) and PIQA (dataset card: unknown) |
text_nota |
40,000 | "none of the other options" questions built by us from text rows | each row follows its parent source; 8,959 rows come from SIQA (no licence found) |
Used as published, not here
These questions were trained on as their sources publish them, with the source's own question and options, so they are not repeated here:
| source | questions (stage 1 and stage 2) | where |
|---|---|---|
| GQA, its own yes/no questions | 306,979 | GQA train_balanced, read from lmms-lab-encoder/GQA |
| SAT | 139,632 | array/SAT-v2 |
| SIQA | 33,377 | allenai/social_i_qa |
| PIQA | 16,113 | ybisk/piqa |
| AI2D | 5,777 | The Cauldron, ai2d |
| ChartQA, its own yes/no questions | 557 | The Cauldron, chartqa |
Overlap with evaluation sets
Rows. No evaluation row is here.
Images. Every training image was compared with the images of nine evaluation sets: our stage-1 and stage-2 dev splits, our DriveLM test split, SAT dev, val and test, CV-Bench, BLINK and 3DSRBench. This includes the images of the questions that are not here. No training image is the same file as an evaluation image.
Near-identical images. Candidate pairs are within 10 bits of each other on a 63-bit DCT hash. A pair counts as the same picture when the aspect ratios agree within 3 % and the mean absolute difference at 64 × 64 grey is at most 6 of 255. On the public benchmarks, the matches are photographs that GQA and TallyQA share with the benchmark, and no benchmark question is asked about them in training:
evaluation set images images with a same-picture training image of those, questions also asked on that picture in training SAT val 435 0 0 SAT test 77 16 (GQA 13, TallyQA 6) 0 CV-Bench 2,638 42 (GQA 42, TallyQA 1) 0 BLINK 509 175 (GQA 121, TallyQA 74) 0 3DSRBench 3,603 1 (TallyQA 1) 0 Our own evaluation splits (not released).
- SAT dev has 2,000 questions on 262 images. 249 of its images have a near-identical SAT training image, but none of its questions is asked on that image in training.
- Our DriveLM test split has 6,571 questions on 2,798 images from 69 scenes of DriveLM-nuScenes' train file. Training uses 558 other scenes, and the dev split uses 69 more.
- Even so, 176 of the test images have a near-identical training image. Each such pair comes from the same nuScenes log, at most 60 s apart, so the two scenes are parts of one drive. On those images, 323 test questions are also asked in training, 276 of them with the same answer.
Text. The text questions were compared with 45,008 questions from ARC-Challenge, ARC-Easy, BoolQ, HellaSwag, MMLU, MMLU-Pro, TruthfulQA (MC1) and WinoGrande. No question shares half or more of its 10-word shingles with a benchmark question. Three
text_yesnoquestions built from OpenBookQA share the question text "Which of the following is true?" with an MMLU-Pro question, with different options.Source splits. Most sources use their train split only. The exceptions:
openimages_boxtakes its images from the OpenImages validation (4,472 questions) and test (13,494) splits.openimages_verticaltakes them from train (1,458), validation (270) and test (760).- SpatialSense uses its train and validation images, and its test split is not used.
- Downloads last month
- 82