---
language: en
license: other
task_categories:
- image-classification
pretty_name: Stickers Binary v2 (Cleaned)
---

# Stickers Binary v2 — Cleaned

Binary SFW/NSFW sticker classification dataset. This version has been cleaned
of likely label errors using cross-validated out-of-fold model predictions
combined with [cleanlab](https://github.com/cleanlab/cleanlab)'s
`find_label_issues`.

## Structure

This dataset has exactly two columns:

| Column  | Type  | Description                                   |
|---------|-------|------------------------------------------------|
| `image` | image | The sticker image, 256x256, letterboxed (see below). |
| `label` | int64 | `0` = SFW, `1` = NSFW.                        |

## Class distribution

| Split | Count | Percentage |
|-------|-------|------------|
| SFW (`0`)  | 97257 | 71.0% |
| NSFW (`1`) | 39713 | 29.0% |
| **Total**  | **136970** | 100% |

## Preprocessing: letterbox resize — REQUIRED at inference time

Every image in this dataset was resized to 256x256 using **letterbox
resizing**: the image is scaled to fit within 256x256 while preserving its
original aspect ratio, then padded with solid gray (`RGB 114, 114, 114`) to
fill the remaining space. This avoids the distortion of a plain squash-resize
and avoids losing content at the edges the way a center-crop would.

**Any model trained on this dataset must receive the same letterbox
preprocessing on every image at inference time — not a plain resize, not a
center-crop.** If inference uses a different resizing strategy than training
did, the model sees a distribution of inputs it never trained on (different
aspect-ratio handling, different effective content scale and position within
the frame), which will silently degrade accuracy without raising an error.

Reference implementation used to build this dataset:

```python
from PIL import Image

def letterbox_resize(img, target_size=(256, 256), pad_color=(114, 114, 114)):
    img = img.convert("RGB")
    img_ratio = img.width / img.height
    target_ratio = target_size[0] / target_size[1]
    if img_ratio > target_ratio:
        new_width = target_size[0]
        new_height = max(1, int(new_width / img_ratio))
    else:
        new_height = target_size[1]
        new_width = max(1, int(new_height * img_ratio))
    resized_img = img.resize((new_width, new_height), resample=Image.Resampling.LANCZOS)
    padded_img = Image.new("RGB", target_size, color=pad_color)
    paste_x = (target_size[0] - new_width) // 2
    paste_y = (target_size[1] - new_height) // 2
    padded_img.paste(resized_img, (paste_x, paste_y))
    return padded_img
```

Apply this exact function (same `target_size`, same `pad_color`) to any image
before passing it to a model trained on this dataset, whether in evaluation,
a serving pipeline, or downstream fine-tuning.