Forturne/SDS-KoPub-OCR

SDS-KoPub OCR Results & Embeddings

OCR layout parsing results and VL embeddings for the SDS-KoPub-VDR-Benchmark corpus (40,781 Korean public document pages).

Contents

File	Description	Size
ocr_results.jsonl	GLM-OCR structured layout results (regions, markdown, bbox, labels)	40,781 records
parsed_texts.jsonl	Extracted text per page (embedding input)	40,781 records
embeddings/corpus_regions.npy	Region multimodal embeddings (image+caption)	(21052, 2048)
embeddings/region_metadata.jsonl	Region metadata (page_id, caption, label)	—
embeddings/corpus_ocr_text.npy	OCR text embeddings	(40781, 2048)
embeddings/queries.npy	Query embeddings	(600, 2048)
crops.tar.gz	Image/chart region crops	21,052 images

Models Used

• OCR: GLM-OCR (0.9B, layout via PP-DocLayoutV3)
• Embeddings: Qwen3-VL-Embedding-2B-FP8 (2048-dim)

OCR Result Format

Each line in ocr_results.jsonl:

{
  "page_id": "doc_123_page_0",
  "page_idx": 0,
  "regions": [
    {"index": 0, "label": "doc_title", "bbox_2d": [x1, y1, x2, y2], "content": "..."},
    {"index": 1, "label": "table", "bbox_2d": [...], "content": "<table>...</table>"},
    {"index": 2, "label": "image", "bbox_2d": [...], "content": null}
  ],
  "markdown": "# Title\n\n| col1 | col2 |\n...",
  "image_crops": [{"path": "crops/doc_123_page_0_crop_2.jpg", "bbox": [...], "label": "image"}]
}

Usage

import json
import numpy as np
from huggingface_hub import hf_hub_download

# Load OCR results
path = hf_hub_download("Forturne/SDS-KoPub-OCR", "ocr_results.jsonl", repo_type="dataset")
with open(path) as f:
    records = [json.loads(line) for line in f]

# Load embeddings
reg_emb = np.load(hf_hub_download("Forturne/SDS-KoPub-OCR", "embeddings/corpus_regions.npy", repo_type="dataset"))
txt_emb = np.load(hf_hub_download("Forturne/SDS-KoPub-OCR", "embeddings/corpus_ocr_text.npy", repo_type="dataset"))
q_emb = np.load(hf_hub_download("Forturne/SDS-KoPub-OCR", "embeddings/queries.npy", repo_type="dataset"))

# Retrieval: cosine similarity (embeddings are L2-normalized)
scores_text = q_emb @ txt_emb.T    # (num_queries, num_pages)
scores_region = q_emb @ reg_emb.T  # (num_queries, num_regions)

Pipeline

Generated with run_b200_pipeline.py on NVIDIA B200 (192GB).