Dataset Viewer
Auto-converted to Parquet Duplicate
image
imagewidth (px)
1.85k
1.85k
smoldocling_text
stringclasses
5 values
inference_info
stringclasses
1 value
127>16>344>21>THE PROPOSOGRAPHY OF THE NEO-ASYRIAN EMPIRE 440>21>473>27>Pali-ereš <text>15>32>238>65>Pali-ereš ("Pali has desired"); Akk.; msc; wr. mdGIDU-KAM-eS", mdGIDU-KAM, mdGIDU- APIN-eS", mdGIDU-APIN, mdGIDU-KAM, note that the name written mdGIDU-U-PAB (Pali-belu-usur) in the copy ADD 488:3 has been embedded to "...
[{"column_name": "smoldocling_text", "model_id": "ds4sd/SmolDocling-256M-preview", "processing_date": "2025-11-10T20:45:59.888407", "batch_size": 16, "max_tokens": 8192, "gpu_memory_utilization": 0.8, "max_model_len": 8192, "output_format": "markdown", "prompt": "Convert page to Docling.", "script": "smoldocling-ocr.py...
135>13>349>18>THE PROSOOPOGRAPHY OF THE NEO-ASSYRIAN EMPIRE 442>19>474>25>Paikuku <text>17>29>230>47>the Assyrian court SAA 7 58 i (not dated or date lost, but cf. Nabāb-bel-šumati 1.).</text> 17>47>230>56>2. Entry in a list of specimen names: "pa-du-u- <text>17>56>230>64>DINGIR ADD App.3 iii 28. C. Ambos</text> <text...
[{"column_name": "smoldocling_text", "model_id": "ds4sd/SmolDocling-256M-preview", "processing_date": "2025-11-10T20:45:59.888407", "batch_size": 16, "max_tokens": 8192, "gpu_memory_utilization": 0.8, "max_model_len": 8192, "output_format": "markdown", "prompt": "Convert page to Docling.", "script": "smoldocling-ocr.py...
23>23>45>28>Palbuq 146>15>361>20>THE PROPOSOGRAPHY OF THE NEO-ASSYRIAN EMPIRE <text>23>32>246>52>O lran. * P aju-k-a based on Iran. * p aju- "protector, guardian" as in Av. p aiu- ( = Ved. p aju- , also as a proper name) from the root * p a "to protect."</text> 23>52>246>72>1. City-lord of Kilambate (reign of Sargon...
[{"column_name": "smoldocling_text", "model_id": "ds4sd/SmolDocling-256M-preview", "processing_date": "2025-11-10T20:45:59.888407", "batch_size": 16, "max_tokens": 8192, "gpu_memory_utilization": 0.8, "max_model_len": 8192, "output_format": "markdown", "prompt": "Convert page to Docling.", "script": "smoldocling-ocr.py...
14>23>29>27>Padi 145>16>358>20>THE PROPOSOGRAPHY OF THE NEO-ASSYRIE EMPIRE <text>29>32>245>96>Padi (hypocr. based on the root pdy "to ran› som, redeem"); WSem.; masc.; wr. "pa-di-i, once "pi-di-i, ft. pdy; Tallqvist (1918) 178, 301; Grön› dahl (1967) 171, 404; Benz (1972) 389; Lipišk 175a) 129-31; Fales (1979a) 60; Za...
[{"column_name": "smoldocling_text", "model_id": "ds4sd/SmolDocling-256M-preview", "processing_date": "2025-11-10T20:45:59.888407", "batch_size": 16, "max_tokens": 8192, "gpu_memory_utilization": 0.8, "max_model_len": 8192, "output_format": "markdown", "prompt": "Convert page to Docling.", "script": "smoldocling-ocr.py...
<text>19>138>239>151>Pabba'u (mng. unknown); Egypt.?: masc.; wr. pma-aba-ba-a-u.</text> <text>19>151>239>223>Horse keeper(?) of Istar of Arbail, father of the bride Mullissu-hammat, from Assur (reign of Esaraddon): pada-ba-a-u. [Lu]*D-71D-B-ANeS*KUR. RA[$]$a[$]$id és URAU.arra-il gilsa herdaughter Mulissu-hammat, a vo...
[{"column_name": "smoldocling_text", "model_id": "ds4sd/SmolDocling-256M-preview", "processing_date": "2025-11-10T20:45:59.888407", "batch_size": 16, "max_tokens": 8192, "gpu_memory_utilization": 0.8, "max_model_len": 8192, "output_format": "markdown", "prompt": "Convert page to Docling.", "script": "smoldocling-ocr.py...

Document Processing using SmolDocling-256M-preview

This dataset contains structured document extraction from images in shaigordin/pna-pages using SmolDocling.

Processing Details

Configuration

  • Image Column: image
  • Output Column: smoldocling_text
  • Output Format: markdown
  • Dataset Split: train
  • Batch Size: 16
  • Max Model Length: 8,192 tokens
  • Max Output Tokens: 8,192
  • GPU Memory Utilization: 80.0%

Model Information

SmolDocling-256M is an ultra-compact multimodal model that excels at:

  • 💻 Code Recognition - Detects and formats code blocks with proper indentation
  • 🔢 Formula Recognition - Identifies and processes mathematical expressions
  • 📊 Tables & Charts - Extracts structured data from tables and charts
  • 📐 Layout Preservation - Maintains document structure with bounding boxes
  • 🏷️ DocTags Format - Efficient minimal representation for documents
  • Fast Inference - Only 256M parameters for quick processing

Dataset Structure

The dataset contains all original columns plus:

  • smoldocling_text: The extracted markdown from each image
  • inference_info: JSON list tracking all OCR models applied to this dataset

Usage

from datasets import load_dataset
import json



# Load the dataset
dataset = load_dataset("{output_dataset_id}", split="train")

# Access the extracted content
for example in dataset:
    
    print(example['smoldocling_text'])
    
    
    break

# View all OCR models applied to this dataset
inference_info = json.loads(dataset[0]["inference_info"])
for info in inference_info:
    print(f"Column: {info['column_name']} - Model: {info['model_id']}")

Reproduction

This dataset was generated using the uv-scripts/ocr SmolDocling script:

uv run https://huggingface.co/datasets/uv-scripts/ocr/raw/main/smoldocling-ocr.py \
    shaigordin/pna-pages \
    <output-dataset> \
    --image-column image \
    --output-format markdown \
    --batch-size 16 \
    --max-model-len 8192 \
    --max-tokens 8192 \
    --gpu-memory-utilization 0.8

Performance

  • Processing Speed: ~0.0 images/second
  • Model Size: 256M parameters (ultra-compact)
  • GPU Configuration: vLLM with 80% GPU memory utilization

Generated with 🤖 UV Scripts

Downloads last month
12