---
license: apache-2.0
language:
- ar
base_model: google/embeddinggemma-2
library_name: transformers
pipeline_tag: feature-extraction
tags:
- arabic
- ocr
- document-ai
- word-matching
- multimodal-embedding
- embeddinggemma
datasets:
- seemorg/books-ocr
---

# ArabicWordsEmbedding

- Fine-tuned from [google/embeddinggemma-2](https://huggingface.co/google/embeddinggemma-2) (440M parameters with the
  audio encoder removed).

## Training

- **Real data:** about 400k word crops from 3,000 scanned pages of classical Arabic printed books
  ([seemorg/books-ocr](https://huggingface.co/datasets/seemorg/books-ocr), human-reviewed transcriptions). Word boxes come from
  Tesseract and are aligned to the transcription.
- **Synthetic data:** generated on the fly in about 50 fonts. Each word is rendered with full, partial or no tashkeel, attached
  punctuation and neighbouring words, then cropped like a word detector would and degraded like a scan (low resolution, blur,
  JPEG, binarisation, noise, rotation).
- **Objective:** contrastive (InfoNCE). Every batch contains lookalike words (same letters without dots), plus targeted pairs
  where tashkeel or punctuation imitates a dot.
- **Method:** LoRA on the vision tower and the transformer, merged into this checkpoint.

## Limitations

- Trained and tested on **printed** Arabic book pages. Handwriting, scene text and decorative fonts were not evaluated.
- Real-data boxes came from Tesseract. A detector that cuts boxes very differently (much tighter, or several words per box)
  may need a quick check.
- Short words that differ by one prefix letter (ل / ال / و) cause most of the remaining real errors.
- Matching inside one page is the intended use. Open-vocabulary recognition is weaker: 93.6% top-1 among 31.7k words.

## License

Apache 2.0, the same as the base model [google/embeddinggemma-2](https://huggingface.co/google/embeddinggemma-2).
Training pages come from [seemorg/books-ocr](https://huggingface.co/datasets/seemorg/books-ocr) (MIT).