H20 CoVT Reproduction Prep

This folder is an outer wrapper around the public CoVT repository. Do not edit upstream/CoVT when preparing or launching the paper-faithful reproduction.

Remote project root:

/data/home/user-test/SHC/COVT

Public code:

upstream/CoVT
remote: https://github.com/Wakals/CoVT.git
commit: recorded in checks/upstream_covt_commit.txt

Local policy for H20:

Workspace is for code, files, lightweight CPU checks, and job submission only.
No direct GPU training/inference in workspace.
No long-running model/data download in workspace.
All future GPU or long-download work must use the prepared jobs/*.sh scripts through rest-sbatch.
Do not read or copy Slurm JWT tokens.
Do not change server system software, drivers, CUDA, or hardware settings.
User-level Python packages may be installed only under the user/project scope.

Reproduction target:

CoVT 3-expert visual-token model: ["sam", "depth", "dino"].
Stage 1 understanding: 6000 optimizer steps.
Public-code Stage 2/3/4 checkpoint: 10000 optimizer steps.
Paper Table-6 3-type checkpoint: 11000 optimizer steps, also prepared for evaluation because the paper describes Stage 4 as 5000 steps.
Evaluation: CV-Bench-2D and CV-Bench-3D through CoVT/VLMEvalKit.

Important distinction:

upstream/CoVT remains public source code.
scripts/ and jobs/ only provide paths, Slurm wrappers, manifests, and verification around the public source.