p4ulbr4dl3y/doc2seek-distill-ocr

Doc2Seek distill OCR: raw page reads

The raw output of Baidu Unlimited-OCR 2 (GGUF, Q4_K_M) over the PDFs in p4ulbr4dl3y/doc2seek-distill-corpus.

Nothing here is parsed, cleaned or chunked. One JSON line is one page, exactly as the decoder emitted it, plus why it stopped and what it cost. Parsing, chunking and any normalization happen offline against the same codec the service uses, so the markup stays recoverable: re-reading a page must never mean a second GPU pass.

Layout

runs/<run-id>/
├── run.json                     # everything the bytes depend on
└── pages/<category>__<host>__<slug>.jsonl

run.json pins the build: text and projector GGUF with their sha256, llama-server version, prompt family, page marker and prompt, dpi, rasterization cap, token budget and ceiling, the no-repeat-ngram guard, context size and the harness version. A second quant, a different prompt or a wider loop guard is a new run directory, never an edit of an existing one — otherwise two answers cannot be compared.

run.json also carries fallback_side, a flag the harness accepts and no longer reads: the raster is now sized from the page's own geometry (pdfinfo -> min(dpi, max_side * 72 / longest_pt)), so the value is a vestige of the earlier guess-and-retry renderer, kept only because a run manifest must not be rewritten after the fact. It was 1600 and it was not used.

One record:

{
  "doc": "finance__www-cbr-ru__svst01122015-4197fd9e.pdf",
  "doc_sha256": "…",
  "category": "finance",
  "source_url": "https://www.cbr.ru/…pdf",
  "page": 4,
  "pages": 9,
  "px": [1654, 2338],
  "raw": "<PAGE><BLOCK>title<LEVEL1>[293, 303, 722, 320]<FORK>…",
  "finish_reason": "stop",
  "tokens_in": 909,
  "tokens_out": 845,
  "elapsed_ms": 21400,
  "chars": 2368,
  "attempts": [{"attempt": 1, "budget": 8192, "finish_reason": "stop", "elapsed_ms": 21400}]
}

attempts is the page's retry ladder, kept in full: a page that stopped at the token budget is answered again on a smaller image and then at a larger budget, and a reader that only keeps the winner loses the evidence that the decoder ran away.

Reproducing the run

harness/ holds the whole control plane, so the run can be repeated rather than described:

file	role
ocr_pages.py	renders one page at a time, calls the server, writes one JSONL line per page
boot.py	entry point for a fresh runtime: fetches the harness over HTTPS and runs it
resume.sh	rebuilds /content from the Hub (models, PDFs, binaries), always re-downloads the harness and prints its sha256
watchdog.sh	keeps a runtime alive: pidfile liveness, three-state job list, account rotation with a 30 min cooldown
progress.sh	one screen of "how far along is it", computed from the uploaded JSONL lines
finish.sh	after FULL_DONE: dedupe, download, chunk, audit
build.sh	compiles the patched llama.cpp on the VM (the alternative to the prebuilt tarball)

Two decisions in there are worth reading before changing anything: liveness is a pidfile, never a process name (a probe that greps for ocr_pages.py matches the colabx bridge that is running the probe), and an unreadable job list means do nothing (a second harness appending to the same files is worse than an idle minute).

The two corpora involved are different things and are easy to confuse: this repo holds the reading of 100 public Russian documents (2569 pages) used to build the chunk corpus, while uocr2-ru-imatrix-calib holds the 92 calibration pages that shaped the quant.

The markup

Unlimited-OCR 2 answers in its own vocabulary:

<PAGE><BLOCK>category<LEVELn>[x1, y1, x2, y2]<FORK>content

<PAGE> opens the page, <LEVELn> states a heading's depth, <FORK> separates the geometry from the content, and tables arrive as <table><tr><td> markup rather than as text lines. All of those markers are structure, never content.

Reading it back

import json
from ml_service.backends.grounding import parse_blocks   # doc2seek, apps/ml

for line in open("runs/<run-id>/pages/<doc>.jsonl"):
    record = json.loads(line)
    blocks = parse_blocks(record["raw"], tuple(record["px"]))

Provenance and licensing

Documents: see the corpus card — public publications, rights with their publishers, official state documents not subject to copyright under Art. 1259(6) of the Russian Civil Code. This dataset holds only the model's reading of them.

Model: baidu/Unlimited-OCR-2 (MIT), quantized to Q4_K_M with a Russian-document imatrix (p4ulbr4dl3y/unlimited-ocr-2-ru-gguf), served by a patched llama.cpp — both patches are in that repo under llama.cpp-patch/.