EunsuKim/mix-culture-food-train

Mix Culture Food Dataset — Train Split

This repository contains the training split for the mix-culture food fine-tuning project. It includes:

• —train_dataset_hf.jsonl: 5,768 multimodal training samples.
• —images/: Only the images referenced in the JSONL, organized by their source dataset.

Each JSONL entry follows the Swift "messages + images" schema. Questions mention either the food itself or the left food (for multi-dish compositions). Answers are deterministic lookups from food_combinations_with_replacements_final_0914.csv.

Directory Layout

.
├── README.md
├── train_dataset_hf.jsonl
└── images/
    ├── 1_white_bg_selected_final_0914/
    ├── 2_synt_final_0914/
    ├── 5_synt_landmark_single/
    ├── 5_synt_street_single/
    ├── 5_synt_landmark_multi/
    └── 5_synt_street_multi/

Usage Example (ms-swift)

swift train \
  --dataset https://huggingface.co/datasets/EunsuKim/mix-culture-food-train \
  --columns messages images

Notes

• —Train split retains all 2_synt_final_0914 combinations to preserve food2 diversity.
• —The other synthetic subsets are down-sampled to roughly one third per base ID while maintaining balanced background coverage.
• —License and redistribution terms should be specified before public release.