meddies-documents-raw

Raw model output and pre-repair snapshots for the HoangHa/meddies-documents corpus. Private archive; exists so converted/repaired data is never the only copy.

Coverage — read this first

Raw Vertex/Gemini output survives ONLY for the 2026-07-17/18 textbook trial runs (6 batches, 5,535 page predictions, ~6 book stems: ASPEN nutrition, Nelson's antimicrobial, ATLS, AAP neonatal resuscitation, WHO maternal health). Each earlier OCR run used a per-run GCS bucket that was deleted after processing, so the main 977-document corpus's raw output is unrecoverable. This gap is why 36 pages across 29 documents could not be repaired in July 2026. Going forward the pipeline archives raw output here as it converts (write-path added before the full textbook round).

Layout

• runs/<run-ts>/<batch-ts>/predictions.jsonl — verbatim Vertex batch prediction output (request + response per page).
• snapshots/bak-2026-07-31-converter-repair/<stem>_data.json.bak — the pre-repair _data.json of the 383 documents rewritten by the 2026-07-31 converter repair. Nearest surviving pre-transformation state for those documents.

Provenance

GCS source buckets: meddies-vertex-ocr-099a99c75f00-{20260717t203057z,20260717t235404z,20260718t092532z}. Corpus context: meddies-document-preprocessing-context ledger 2026-08-20.