Skip to content

Latest commit

 

History

History

README.md

Pre-Training Workspace

Folder structure

  • DATASET/: Input PDFs (example: Release_1/).
  • outputs/[timestamp]_[dataset]/: everything produced from one PDF dataset, self-contained in a single folder:
    • [slug]-page-[n].png — the converted pages.
    • [timestamp]_[dataset]-OCR.csv — OCR text per page (Surya OCR).
    • [timestamp]_[dataset]-SUMMARIES.csv — per-page summary (local Gemma 3).
    • [timestamp]_[dataset]-LAYOUT.csv — per-page layout description, image-grounded (local Gemma 3).
    • [timestamp]_[dataset]-QA.csv — synthetic QA pairs per page, one row per pair (local Gemma 3).

Keeping the PNGs and their CSVs in the same timestamped folder means each outputs/[timestamp]_[dataset]/ is a complete, portable unit for that run, and re-running a step against the same folder resumes instead of colliding with a different run of a similarly-named dataset.

outputs/ is where all generated PNGs and CSVs go, and none of it is committed to git — only outputs/README.md is tracked; everything else in that folder is git-ignored.

Run Files

uv_setup.bat     :: create/sync local venv, install CUDA torch, validate CUDA
exec_1.bat       :: Step 1 - Convert PDF dataset to PNG pages (resumable)
exec_2.bat       :: Step 2 - OCR PNG pages with Surya OCR (resumable)
exec_3.bat       :: Step 3 - Summarize OCR with local Gemma 3 (resumable)
exec_4.bat       :: Step 4 - Describe page layout from PNGs with local Gemma 3 (resumable)
exec_5.bat       :: Step 5 - Generate synthetic QA pairs with local Gemma 3 (resumable)
main.bat         :: interactive menu covering all pipeline steps

Each exec_N.bat is a standalone, double-clickable entry point for one pipeline step: it bootstraps the env via uv_setup.bat, then runs the matching script under scripts/. main.bat still covers the full menu, including steps that don't have an exec_N.bat yet.

All operational batch and Python scripts are now under scripts/.