Legend2727/xLingual-picobanana-taxonomy-6k

---
dataset_info:
dataset_name: xLingual PicoBanana Taxonomy 6k
license: mit
task_categories:
- image-editing
- multimodal-classification
language:
- en
- hi
- bn
pretty_name: xLingual PicoBanana Taxonomy 6k
---

# xLingual PicoBanana Taxonomy 6k

Canonical public dataset repo:

- `Legend2727/xLingual-picobanana-taxonomy-6k`

This release is the cleaned 6k subset of the earlier 12k repository. It keeps:

- source image
- edited image
- instruction in English, Hindi, and Bangla
- canonical 11-label taxonomy labels

## Dataset contract

- rows: 6000
- languages: en / hi / bn
- taxonomy labels: 11 canonical classes
- source_type distribution: `{"preference_rejected": 3893, "sft": 2107}`

## Files

- `metadata.jsonl`: full 6k row-level metadata
- `labels.jsonl`: label-focused companion file
- `images/source/`: source images
- `images/target/`: edited images
- `splits/`: train/val/test/demo split files

## Loading

Use the repo as a standard JSONL + image-path dataset. The metadata rows contain POSIX paths and all three prompt languages.

## Notes

- The older 12k repository remains public only for historical compatibility.
- This 6k repo is the preferred public release for downstream work.