---
license: apache-2.0
task_categories:
  - image-classification
language:
  - hi
  - mr
  - ne
  - sa
  - bn
  - as
  - pa
  - gu
  - or
  - ta
  - te
  - kn
  - ml
  - ur
  - sat
  - mni
  - en
tags:
  - font-identification
  - font-recognition
  - indic
  - typography
  - synthetic
pretty_name: Lipika eval — Indic font recognition benchmark
size_categories:
  - 1K<n<10K
dataset_info:
  features:
  - name: image
    dtype: image
  - name: family
    dtype: string
  - name: script
    dtype: string
  - name: weight
    dtype: int32
  - name: italic
    dtype: bool
  - name: font_file
    dtype: string
  - name: font_license
    dtype: string
  - name: font_source
    dtype: string
  splits:
  - name: validation
    num_bytes: 293259746
    num_examples: 6876
  download_size: 289898553
  dataset_size: 293259746
configs:
- config_name: default
  data_files:
  - split: validation
    path: data/validation-*
---

# Lipika eval — Indic font recognition benchmark

The frozen validation set behind **[loopdesk-ai/lipika](https://huggingface.co/loopdesk-ai/lipika)**
(Indic font recognizer): **6,876 synthetic text crops** covering **553 freely-licensed font
families** across **13 scripts** (Devanagari, Bengali, Gujarati, Gurmukhi, Kannada, Malayalam,
Meetei Mayek, Odia, Ol Chiki, Perso-Arabic, Tamil, Telugu, Latin).

This is the set reported as "synthetic val" in the model card (Lipika v2.4 scores **0.849
family top-1 / 0.977 top-5 / 0.991 script**). Use it to benchmark your own font-ID model or
to reproduce ours.

## How it was made

Crops are rendered with proper Indic shaping (HarfBuzz via a Pillow/raqm stack) from the
open [Lipika font corpus](https://github.com/Loopdesk-AI/lipika), with the frozen-eval
augmentation distribution: varied sizes, colors, backgrounds, mild blur/noise/rotation,
JPEG artifacts. Seed-frozen (`seed 2028`) — the generator lives in
[`fontrecog/dataset/pregenerate_val.py`](https://github.com/Loopdesk-AI/lipika/blob/main/src/fontrecog/dataset/pregenerate_val.py).

**Training data is not published as files** — it is rendered on the fly from font files by
the same pipeline; the corpus manifest + fetch scripts in the GitHub repo reproduce it.

## Fields

| Field | Description |
|---|---|
| `image` | RGB text crop |
| `family` | Font family label (one of 553) |
| `script` | Writing script of the rendered text |
| `weight` | CSS weight bucket 100–900 |
| `italic` | Italic/oblique flag |
| `font_file` | Source font filename |
| `font_license` | Upstream font license (OFL, GPL+FE, freeware, …) |
| `font_source` | Corpus source (google_fonts, smc, ekushey, salrc, …) |

## Licensing

Dataset packaging & annotations: **Apache-2.0**. Images are bitmap renders of freely
licensed / freely distributed fonts only (OFL, GPL with font exception, public freeware);
each row carries its font's upstream license in `font_license`. Crops rendered from
proprietary system fonts (10 families, 84 crops of the original 6,960) are **excluded** from
this release; scores on this subset may differ from the model card's full-set numbers by
<0.2pt.

## Load

```python
from datasets import load_dataset
ds = load_dataset("loopdesk-ai/lipika-eval", split="validation")
```

## Citation

```bibtex
@software{lipika2026,
  title  = {Lipika: Indic Font Recognition},
  author = {Loopdesk},
  year   = {2026},
  url    = {https://huggingface.co/loopdesk-ai/lipika}
}
```