VeraPhys Camera-Ready Code

This repository contains the source code and scripts needed to run the VeraPhys pipeline. It intentionally excludes result artifacts, response logs, prediction files, benchmark labels, model caches, and local credentials.

Setup

From this directory:

uv sync
source .venv/bin/activate

Alternatively:

python -m venv .venv
source .venv/bin/activate
pip install -e .

Configure endpoints either through environment variables or a local .env copied from .env.example:

cp .env.example .env
# edit OPENAI_BASE_URL / MODEL / PADDLEOCR_VL_BASE_URL as needed

The main VLM and PaddleOCR-VL servers are expected to expose OpenAI-compatible /v1/chat/completions APIs.

Run VeraPhys on SeePhysPro

ARTIFACT_DIR=artifacts/veraphys_full \
REUSE_TESTMINI=/path/to/testmini_predictions.jsonl \
PARALLEL=64 \
SELECTOR_PARALLEL=32 \
bash scripts/seephyspro-currentbest-full.sh

Important variables:

• DATASET: defaults to Kun-Xiang/SeePhysPro.
• MODEL: defaults to Qwen/Qwen3.6-35B-A3B.
• BASE_URL or OPENAI_BASE_URL: main VLM endpoint.
• PADDLEOCR_VL_BASE_URL: PaddleOCR-VL endpoint.
• REUSE_TESTMINI: public test-mini prediction JSONL to append for the all-with-testmini package. This is not bundled here.

Main outputs:

• ${ARTIFACT_DIR}/predictions-seephyspro-test.jsonl
• ${ARTIFACT_DIR}/submission-seephyspro-test.zip
• ${ARTIFACT_DIR}/predictions-seephyspro-all-with-testmini.jsonl
• ${ARTIFACT_DIR}/submission-seephyspro-all-with-testmini.zip

Run the OCR-VL Enhancement Layer

For an existing base prediction file:

BASE=/path/to/base_predictions.jsonl \
ARTIFACT_DIR=artifacts/ocrvl_layer \
PARALLEL=32 \
SELECTOR_PARALLEL=32 \
bash scripts/paddleocr-vl-enhance.sh

Useful Utilities

Postprocess predictions:

python -m track3_cli.postprocess_predictions \
  --input raw_predictions.jsonl \
  --output postprocessed_predictions.jsonl

Package a submission:

python -m track3_cli.package_submission \
  --predictions predictions.jsonl \
  --output submission.zip