arxiv_chart_qa_02/README.md

arxiv_chart_qa_02

Download each paper’s source archive from the image_sources[].source_url URL under the paper’s terms. For the full collection, use arXiv bulk source data; bulk_source_member identifies the paper archive inside the monthly bulk tar. Use the paper source archive, rather than the paper PDF. Legacy arXiv IDs retain their slash in the URL.

For each annotation, image_sources gives the source paper (arxiv_id), the exact figure filename inside its archive (source_path), and the output path (image). Write the reconstructed figure directly to that output path relative to the release repository root; it matches the message’s image reference. If source_path_encoding is percent-encoded-bytes, percent-decode source_path to bytes when looking up the archive member, and retain the encoded spelling for the output filename.

Apply the method specified by render_method:

• copy: copy original PNG/JPEG bytes unchanged.
• pdf: render the first page of the source figure PDF with PyMuPDF using page.get_pixmap(dpi=144, alpha=False), and save as PNG.
• postscript: render EPS/PS with Ghostscript using the command below. It renders at 288 dpi, downsamples by two to 144 dpi, crops to the EPS bounding box, and uses a white background.
• raster_to_png: open the source with Pillow, select the first frame, preserve RGB/RGBA (convert other modes to RGB, or RGBA if alpha is present), and save as PNG without resizing.

gs -dSAFER -dBATCH -dNOPAUSE -dTextAlphaBits=4 -dGraphicsAlphaBits=4 \
  -dDOINTERPOLATE -sDEVICE=png16m -r288 -dEPSCrop -dDownScaleFactor=2 \
  -sOutputFile="OUTPUT_PATH_FROM_JSONL" "SOURCE_FIGURE.eps"

Install PyMuPDF, Pillow, and Ghostscript for the methods you use. For PostScript figures with external fonts, supply the font directory with Ghostscript’s -I option. Create the output parent directories before writing files.