w4ashabii/nepali_crop_data

Nepali Crop Disease Dataset

A consolidated image dataset for crop identification and plant disease detection, merged from multiple public sources and organized by crop, with an accompanying ShareGPT-format conversational dataset for vision-language model finetuning.

Dataset Summary

Images are grouped one folder per crop (e.g. tomato/, apple/, maize/), each labeled with both a crop and a disease (or healthy). The corpus is class-balanced by design:

• Crops with fewer than 300 images were dropped entirely (9 crops kept out of 17 detected in the source pool).
• Every kept crop is capped at the same per-class quota (equal counts across crops — no crop dominates the dataset).
• Within each crop, images are sampled round-robin across disease labels, so a single disease can't dominate that crop's folder.
• Final split per crop: 4,443 train / 666 val / 666 test, for 47,489 images total across 9 crops.

Labels come from two sources, tracked per-image via a label_source field:

• metadata — parsed from existing .jsonl/.csv/.json annotation files shipped with a source dataset (higher precision).
• foldername — inferred by keyword-matching the original folder/file path when no structured metadata was available.

Source Datasets

This dataset merges images from:

Source	Link
Kaggle — Crop Disease Detection Dataset	https://www.kaggle.com/datasets/snikhilrao/crop-disease-detection-dataset
GitHub — PlantDoc-Dataset	https://github.com/pratikkayal/PlantDoc-Dataset
Hugging Face — bd-crop-vegetable-plant-disease-dataset	https://huggingface.co/datasets/Saon110/bd-crop-vegetable-plant-disease-dataset

Mendeley Data (6243z8r6t6) was originally scoped for inclusion but was dropped from this build.

Please refer to each original source for its own license and citation requirements — this dataset does not override or supersede those terms. If you are a maintainer of one of the source datasets and have concerns about inclusion or attribution, please open a discussion on this repo.

Dataset Structure

train/<crop>/*.jpg
val/<crop>/*.jpg
test/<crop>/*.jpg
labels_registry.jsonl
sharegpt.jsonl
zips/dataset_part_*.zip   # full archive, for bulk download

labels_registry.jsonl

One row per image:

{"image": "train/tomato/abc123.jpg", "split": "train", "crop": "tomato", "disease": "early blight", "label_source": "metadata"}

sharegpt.jsonl

One conversation per image, formatted for VLM instruction finetuning:

{
  "id": "0000001",
  "image": "train/tomato/abc123.jpg",
  "conversations": [
    {"from": "human", "value": "<image>\nWhat crop is shown in this image, and does it have any disease? If so, name the disease."},
    {"from": "gpt", "value": "This is a Tomato leaf affected by early blight."}
  ]
}

Healthy images get an explicit no-disease answer; images where the disease label couldn't be confidently determined are marked as such rather than guessed.

How to Load

As an image folder (classification)

from datasets import load_dataset

ds = load_dataset(
    "imagefolder",
    data_dir="train",  # or point at an extracted local copy
)

As ShareGPT conversations (VLM finetuning)

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="sharegpt.jsonl",
)

Note: image paths in sharegpt.jsonl are relative to the repo root (e.g. train/tomato/abc123.jpg) — resolve them against your local copy of the train/val/test folders (or the extracted zips/ archive) before loading images.

From the zipped archive

from huggingface_hub import hf_hub_download
import zipfile

local_zip = hf_hub_download(
    repo_id="w4ashabii/nepali_crop_data",
    repo_type="dataset",
    filename="zips/dataset_part_000.zip",
)
with zipfile.ZipFile(local_zip) as zf:
    zf.extractall("training_data")

Known Limitations

• Crop and disease labels are extracted via keyword matching where source metadata wasn't available (label_source: "foldername") — these may contain classification errors and should be spot-checked before use in high-stakes applications.
• Crops with fewer than 300 available images were dropped to keep the balancing meaningful — this dataset does not cover every crop present in the original source datasets, only the 9 that had enough images.
• Balancing is enforced at the crop level (equal images per crop) and approximately at the disease level within each crop (round-robin sampling) — it is not a strict guarantee of equal counts per individual disease across the whole dataset. Check labels_registry.jsonl if you need exact per-disease counts.
• Image quality, resolution, and photographic conditions vary across the source datasets.

Citation

If you use this dataset, please cite the original source datasets listed above in addition to this repository.