SonexaAI/Sonexa-Image-1-Data

A bilingual (English + Russian) image-caption dataset derived from ai-enthusiasm-community/CC3M-35L. This is a curated subset intended for image-text pretraining and multimodal fine-tuning where only English and Russian captions are required.

Source

• Original dataset: ai-enthusiasm-community/CC3M-35L
• Underlying corpus: Google Conceptual Captions 3M (CC3M), augmented with multilingual captions via the Crossmodal-3600 methodology (Thapliyal et al., 2022).
• Original languages: 35 languages, with one caption_<lang> list-column per language per row, plus the raw image (decoded bytes) and image_uid.

What this dataset contains

This repository keeps only the columns relevant for English + Russian image-text work:

Field	Type	Description
image_uid	string	Original CC3M image identifier.
caption_uid	list[string]	Per-caption UIDs in the form {image_uid}_{comment_number}.
image	image	Decoded image (bytes + path), HuggingFace Image feature.
caption_en	list[string]	One or more English captions for the image.
caption_ru	list[string]	One or more Russian captions for the image.

All 33 other caption_<lang> columns from the source were dropped.

Subset construction

The original CC3M-35L dataset is ~266 GB across 267 train parquet shards (2,855,284 examples) and 2 validation shards (13,443 examples). To produce a tractable ~6 GB subset suitable for experimentation:

1. Train split: the first 5 of 267 shards were taken (train-00000 through train-00004), yielding 53,470 examples.
2. Validation split: both validation shards were taken in full (validation-00000 and validation-00001), yielding 13,443 examples.
3. For each shard, only the 5 columns listed above were retained and the result was re-encoded as Parquet (zstd compression, schema preserved).

No row-level filtering was applied: every row in the selected shards is kept as-is. Other-language captions are simply not part of the output schema.

Splits and sizes

Split	Files	Examples	Size (GiB)
train	5	53,470	4.42
validation	2	13,443	1.01
Total	7	66,913	5.43

Usage

from datasets import load_dataset

ds = load_dataset("SonexaAI/Sonexa-Image-1-Data")
print(ds)
# DatasetDict({
#     train:      Dataset({ features: [...], num_rows: 53470 })
#     validation: Dataset({ features: [...], num_rows: 13443 })
# })

example = ds["train"][0]
print(example["image_uid"])      # '000000000001'
print(example["caption_en"])     # ['a very typical bus station']
print(example["caption_ru"])     # ['Очень типичная автобусная станция']
example["image"]                 # PIL.Image.Image

Streaming is also supported:

ds = load_dataset("SonexaAI/Sonexa-Image-1-Data", streaming=True)
for ex in ds["train"]:
    ...

Licensing

• The source dataset ai-enthusiasm-community/CC3M-35L is a derivative of Google's CC3M and the Crossmodal-3600 multilingual extension. CC3M images and captions are released under their respective original licenses (see the CC3M release notes). The Crossmodal-3600 annotations are released for research use.
• This derivative subset inherits the same licensing constraints. Users must comply with the original CC3M terms and the Crossmodal-3600 terms when using, redistributing, or building upon this data.
• Commercial use of the underlying images is not automatically granted; please consult the original CC3M terms before any commercial use.

Citation

If you use this dataset, please cite both the original source and Crossmodal-3600:

@misc{thapliyal2022crossmodal,
  title   = {Crossmodal-3600: A Massively Multilingual Multimodal Evaluation Dataset},
  author  = {Thapliyal, Ashish V. and Pont-Tuset, Jordi and Chen, Xi and Soricut, Radu},
  year    = {2022},
  eprint  = {2205.12522},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

@inproceedings{sharma2018cc3m,
  title     = {Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset for Automatic Image Captioning},
  author    = {Sharma, Piyush and Ding, Nan and Goodman, Sebastian and Soricut, Radu},
  booktitle = {ACL},
  year      = {2018}
}

Provenance

• Curated by: SonexaAI
• Date of preparation: 2026-10-17
• Source dataset: ai-enthusiasm-community/CC3M-35L
• Filtering script: scripts/filter_cc3m.py (run end-to-end, resumable).