README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.jsonl.gz
tags:
  - pdf
  - ocr
  - chandra
  - chandra-ocr-2
  - markdown
  - html
  - hf-jobs
  - uv-script

PDF OCR with Chandra OCR 2

This output bundle stores OCR results for PDFs referenced by a supplied URL list using datalab-to/chandra-ocr-2.

Summary

• Output bucket: hf://buckets/sroecker/pdf-chandra-ocr
• Source PDF URLs in input list: 111
• Processed inputs recorded in state/processed_inputs.txt: 111
• Successes: 111
• Partial successes: 0
• Errors: 0
• Next shard index: 12
• Updated at: 2026-04-15T18:56:12.558178+00:00

Files

• data/part-*.jsonl.gz: OCR result shards, one JSON object per PDF
• state/processed_inputs.txt: completed PDF URLs used for resume
• state/summary.json: aggregate counters and bookkeeping

Each record includes:

• num_pages: total number of pages in the source PDF
• num_pages_processed: number of pages actually sent to OCR
• pdf_exceeds_page_limit: whether the PDF had more pages than the configured OCR cap
• max_pages_per_paper: configured OCR page cap for the run

Load the results

from datasets import load_dataset

dataset = load_dataset("<dataset-id>", data_files="data/*.jsonl.gz", split="train")
print(dataset[0]["source_id"])
print(dataset[0]["pdf_url"])
print(dataset[0]["markdown"][:1000])

Job config

• Prompt type: ocr_layout
• Page batch size: 28
• Max output tokens: 12384
• Max model length: 18000
• GPU memory utilization: 0.85
• Minimum download request interval: 0.0 seconds
• Max pages per paper sent to OCR: 30
• Bucket backend: hf-cli
• Paginate output: False
• Include headers/footers: False

Reproduction

hf jobs uv run --flavor l4x1 --image vllm/vllm-openai:v0.17.0 \
  -s HF_TOKEN --timeout 2d \
  ./chandra2-arxiv-ocr.py --output-dataset hf://buckets/sroecker/pdf-chandra-ocr \
  --output-bucket hf://buckets/sroecker/pdf-chandra-ocr \
  --pdf-urls-url https://.../pdf_urls.txt