Add submission for glgh - #79
Open
glgh wants to merge 1 commit into
Open
Conversation
Offline OCR + classical CV + a hand-written rules cascade. No LLM, VLM, cloud API or network at runtime. - predictions.jsonl: all 5,000 validation cases, produced in one uninterrupted container run under the exact contract limits (23,086 s, 4.62 s/PDF against the 6 s budget). - MEMO.md: approach, adversarial handling, failure modes, what I would improve. - SUBMISSION.md: link to the public solution repo (which contains the Dockerfile), contract-compliance table, and a disclosure of every constant derived from the training labels.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Offline document pipeline: PyMuPDF for text and layout, Tesseract for OCR, NumPy/Pillow for scan-repair geometry, and a hand-written 17-branch rules cascade for adjudication. No LLM, VLM, multimodal model, cloud OCR/document API, or network access at runtime. There are no learned model artifacts in the image — the only fitted numbers are a ~2 KB per-branch confidence table checked into the repo as JSON.
Public solution repo (contains the
Dockerfile): https://github.com/glgh/mib-doc-challenge-solutionContents
predictions.jsonlMEMO.mdSUBMISSION.mdRuntime, measured under the exact contract limits
One uninterrupted container run with
--network none --cpus 4 --memory 8g --pids-limit 512 --read-only:One caveat stated plainly in
SUBMISSION.md: this was measured on Apple silicon, and the contract fixes the vCPU count but never the vCPU speed. On an SMT instance "4 vCPU" is two physical cores, which is the bad case for cache-hungry Tesseract. The pipeline streams and flushes each row as it completes, so a container stopped at the limit still leaves a valid, scoreable partial submission.Validation
5,000 rows, 5,000 unique ids, schema-clean against
schemas/submission.schema.json, exact manifest coverage (0 missing, 0 unexpected),scripts/validate_submission.pyclean. On the frozen 700/300 train split: dev 128.24 / 150, holdout 128.28 / 150, 0 catastrophic false approvals, 0 missing rows.Only
submissions/glgh/is touched.🤖 Generated with Claude Code