---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: image
    dtype: image
  - name: output
    dtype: string
  - name: file_name
    dtype: string
  splits:
  - name: train
    num_bytes: 63343704014
    num_examples: 151945
  download_size: 0
  dataset_size: 63343704014
license: other
task_categories:
- image-classification
size_categories:
- 100K<n<1M
license_name: mixed-creative-commons
license_link: https://github.com/inaturalist/inaturalist-open-data
---

# NE Plant Classes

Plant photographs from the [iNaturalist open data](https://github.com/inaturalist/inaturalist-open-data)
archive, filtered to a New England bounding box (latitude 41 to 48, longitude
-74 to -67) and each labelled by a vision language model with the kind of
subject it depicts.

Labels come from [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B);
they are model output, not human annotation, and have not been checked against
a hand-labelled sample.

151,945 images in a single `train` split.

## Columns

- `image` - the photograph, embedded in the dataset
- `output` - the model's one-word classification
- `file_name` - the original image filename, named after the iNaturalist photo id

## Labels

The prompt sorts each photo into one of five classes, falling back to `other`
when none of the first four fit.

![Three example photos for each label](https://raw.githubusercontent.com/belambert/plant-pics/main/images/ne_plant_classes.jpg)

| label     |   count |  share |
|-----------|--------:|-------:|
| nature    | 114,192 |  75.2% |
| human     |  33,702 |  22.2% |
| manmade   |   2,310 |   1.5% |
| magnified |   1,715 |   1.1% |
| other     |      26 |  <0.1% |

## Label Quality

The labels are model output and some of them are wrong. A ViT classifier
trained on this dataset
([blambert/ne_plant_classes_vit](https://huggingface.co/blambert/ne_plant_classes_vit))
disagrees with the label on 1.6% of held-out photos, and a review of 160 of
those disagreements suggested the label, not the classifier, was wrong about
two thirds of the time. That puts **roughly 1% of the dataset, on the order of
1,500 photos, under a wrong label**, and that is a lower bound: the classifier
learned the labeller's habits, so mistakes the two share are invisible to it.

The errors are concentrated:

| label     | disagreement rate | note                                                     |
|-----------|------------------:|----------------------------------------------------------|
| nature    |              0.9% | mostly photos with a hand or a ruler in frame            |
| human     |              1.5% | often no hand visible at all                             |
| magnified |              9.9% |                                                          |
| manmade   |             31.4% | roughly 1 in 6 `manmade` labels looks wrong              |

`manmade` is the least reliable label, partly because the prompt never settled
what it means: a plant growing in a sidewalk crack or beside a road is
sometimes `manmade` and sometimes `nature`.

This is still good enough for some purposes. Separating `nature` from `human`
is reliable enough for coarse filtering, which is what
[blambert/ne_plant_photos](https://huggingface.co/datasets/blambert/ne_plant_photos)
does. Treat the small classes, and any per-label accuracy measured against
these labels, with more caution, and hand-label a sample before trusting them
as ground truth.

## Licensing

The photographs keep whichever license each iNaturalist observer chose, so the
collection is not under a single license and most of it is non-commercial.

| license     |   count |  share |
|-------------|--------:|-------:|
| CC-BY-NC    | 114,639 |  75.4% |
| CC-BY       |  22,883 |  15.1% |
| CC0         |  10,080 |   6.6% |
| CC-BY-NC-SA |   3,292 |   2.2% |
| CC-BY-SA    |   1,051 |   0.7% |

Everything but the CC0 portion requires attribution. Observer and photo
metadata live in the iNaturalist open data archive, joinable on the photo id in
`file_name`. The sample photos on this card are all CC0.

## Provenance

Labelled by running `vlm process` with `prompts/inat_classify.txt` over the
downloaded images, then uploaded with `vlm upload-dataset`. Both commands come
from the [vlm-toolkit](https://github.com/belambert/vlm-toolkit) repo; the
filtering and download steps are in [plant-pics](https://github.com/belambert/plant-pics).