al-lataif-al-musawwara-magazine-ocr

Al-Lataif Al-Musawwara Magazine OCR Pages This dataset contains rendered Arabic magazine page images from Al-Lataif Al-Musawwara paired with page-level text and line-level bounding boxes. It is intended for OCR, document understanding, and VLM fine-tuning experiments. Fine-Tuning Notebook A standalone Google Colab notebook for DeepSeek-OCR 3B + TRL SFT is available here: Open the fine-tuning notebook in Colab The notebook can train on this dataset alone or on all three Arabic magazine OCR datasets. Dataset Structure Each row represents one page image. | Column | Description | | --- | --- | | `image` | Page image rendered as a Hugging Face image column. | | `image_id` | Stable row id derived from the source path. | | `source_path` | Original relative image path. | | `collection` | Collection folder name. | | `volume` | Source volume folder name. | | `page_number` | Page number parsed from the filename. | | `width`, `height` | Image dimensions in pixels. | | `ocr_text` | Full page transcription text. | | `line_count` | Number of detected text lines. | | `line_texts` | Text from each detected line. | | `line_bboxes` | JSON array of line polygon bounding boxes. | | `lines_json` | JSON array containing line text and polygon bounding boxes. | | `instruction`, `response` | Image-to-text fine-tuning pair. The instruction is `Extract all text from this image.` | | `task` | Task label. | Splits Split mode: `volume`. | Split | Rows | Parquet shards | | --- | ---: | ---: | | train | 3846 | 4 | | validation | 32 | 1 | | test | 32 | 1 | Total rows: 3910 Usage Because this dataset is large, streaming is the most convenient way to inspect or train from it: For VLM fine-tuning, use `image` as the visual input, `instruction` as the prompt, and `response` as the target answer. For non-streaming use on machines with enough disk/cache space: Curation Notes Images without matching valid OCR JSON were skipped. Common source/footer marks and angled watermark fragments were filtered from th