katha-ai-iiith/one_token_at_a_time

One Token at a Time — companion data

Companion dataset for katha-ai/one_token_at_a_time, the code release for Attending to Multimodal Generation One Token at a Time. This holds the task CSVs, images, precomputed POS-tag outputs, and derived evaluation results needed to run and reproduce the repo's five supported tasks: Fruit-Math, Fruit-Sport, Math-Fruit, VSR, and ChartQA.

Directory structure

data/
├── blank_black_image.png      (not included — see note below)
├── images/
│   ├── oid_media/              Fruit-Math / Fruit-Sport images (Open Images Dataset)
│   ├── Fruits_Clean_12/        Fruit-Sport images (custom-created)
│   ├── ChartQA/                Chart images (ChartQA benchmark)
│   └── vsr_images/              VSR images (COCO)
├── csv/                         Task CSVs (prompt/ground-truth data) + a few derived accuracy tables
├── tagged_outputs/otat_tagged_outputs/   Precomputed Gemini POS-tag outputs, used by blocking/boosting configs
├── automated_gemini_acc_evals/  Our Gemini-graded accuracy results (Fruit-Math family: vanilla, blocking, boosting, confusion-matrix/error-profiling)
├── bar_plot_csv/, global_means/  Compiled inputs for data_plotting/ notebooks
├── other_gemini_scaling_stuff/  Raw Gemini batch-tagging/eval jsonl (see note on FrMaSc below)
├── spot_check_results/, attention_boosting_results/  Small qualitative/spot-check outputs

How this maps to the code repo

configs/*/*.yaml's data.csv_path / data.image_base_dir and boosting.style.pos_dir fields expect this entire data/ directory placed at the repo root, e.g.:

one_token_at_a_time/
├── data/          <- this dataset, unzipped here
├── configs/
├── run_analysis.py
└── ...

VSR, Fruit-Sport, and Fruit-Math's example blocking.yaml/boosting.yaml configs are all reproducible out of the box — their pos_dirs (vsr/vanilla/<model>/vsr/, FrSc_Sp/<model>/fruit_sport/, FrMaSc/<model>/fruit_math/) are populated and correctly laid out here. (Fruit-Sport and Fruit-Math's source data originally used a flatter/differently-located directory layout than fetch_block_indices.py expects — we relocated it into the correct <model>/<task_name>/ nesting under tagged_outputs/otat_tagged_outputs/ when building this dataset; the files themselves are untouched, only their paths changed.) ChartQA and Math-Fruit ship a vanilla.yaml only, no blocking/boosting example.

blank_black_image.png (referenced by every config's black_image_path) is not included — it's dead config: no code in the shipped repo actually reads that field. Safe to ignore, or point it at any placeholder image if you'd rather not leave it dangling.

License and provenance — read this before using the images

All CSV/tagged-output/eval-result data in this dataset (everything except the four images/ subdirectories) is our own derived work, released under CC BY-NC-SA 4.0 (matching the code repo's license) — use, adapt, and redistribute for non-commercial purposes with attribution, under the same license.

The images/ subdirectories are not uniformly ours and carry different provenance:

• images/oid_media/ — sourced from Google's Open Images Dataset. Individual images are listed as CC BY 2.0. Per OID's own documentation: "while we tried to identify images that are licensed under a Creative Commons Attribution license, we make no representations or warranties regarding the license status of each image and you should verify the license for each image yourself." (OID's annotations, not directly relevant here, are separately CC BY 4.0.)

• images/vsr_images/ — sourced from MS-COCO via the Visual Spatial Reasoning dataset (that repo's own code is Apache-2.0, which covers their code, not the images). These images originate from COCO/Flickr and retain their original photographers' individual licenses. COCO does not grant a blanket redistribution license for its images — this subset is included here only to make this specific research reproducible. These images belong to their original creators, not to us or to this project.

• images/ChartQA/ — sourced from the ChartQA benchmark (GPL-3.0 repo license), whose chart images were originally drawn from public sources including Statista and Pew Research Center. Those source sites may hold their own copyright over the chart images. Included here only to ease reproducibility — these belong to their original creators.

• images/Fruits_Clean_12/ — our own custom-created images. CC BY-NC-SA 4.0, same as the rest of this dataset.

If you are a rights holder for any image in vsr_images/ or ChartQA/ and want it removed from this dataset, please open an issue on the code repo.

A note on content

This is a research artifact, not a moderated dataset. Some source CSVs (e.g. VSR's caption / extended_caption fields) contain free-text descriptions generated as part of the original benchmark's or this paper's pipeline (including LLM-generated text) and have not been separately vetted for content beyond what the original benchmarks' own release processes did.

Citation

@misc{gupta2026attending,
      title={Attending to Multimodal Generation One Token at a Time},
      author={Varun Gupta and Vineet Gandhi and Makarand Tapaswi},
      year={2026},
      eprint={2607.03738},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}