Release

OSCaR is published as a coordinated GitHub + Hugging Face release:

• GitHub code repo: nguyennm1024/OSCaR
• Hugging Face dataset: ali-vosoughi/oscar-dataset
• Hugging Face model namespace: ali-vosoughi

Prepare The Staged Bundle

From the bundle root:

uv venv --python 3.10 .venv
source .venv/bin/activate
uv pip install -e ./release_repos/github/OSCaR[release]
python ./release_repos/github/OSCaR/scripts/release/prepare_hf_release.py \
  --bundle-root /path/to/llava_statechange_hf_release_bundle \
  --namespace ali-vosoughi

That command generates:

• dataset card at release_repos/huggingface/datasets/llava-statechange-dataset/README.md
• normalized metadata tables under metadata/
• model cards in each staged model directory
• release_repos/huggingface/models/upload_manifest.json

Publish To Hugging Face

Authenticate with a token through the environment only:

export HF_TOKEN=...
python ./release_repos/github/OSCaR/scripts/release/publish_hf_release.py \
  --bundle-root /path/to/llava_statechange_hf_release_bundle \
  --namespace ali-vosoughi

Notes:

• adapter and projector repos are uploaded when their local staging folders exist
• merged-model repos are skipped unless the corresponding merged model folders are present
• the token is never written to tracked files

Resumable Dataset Upload

The dataset is large enough that it should be uploaded from tmux with the resumable uploader instead of one monolithic pass.

From the bundle root:

export HF_TOKEN=...
./upload_oscar_dataset_tmux.sh |& tee upload_oscar_dataset.log

If the session is interrupted, rerun the same command. Progress is tracked in:

/path/to/llava_statechange_hf_release_bundle/.hf_dataset_upload_state.json

The uploader commits:

• root README.md
• splits/
• manifests/
• metadata/
• eval/
• annotation folders
• data/object-state-data/ in participant-prefix batches such as P01, P02, P22

Publish The GitHub Repo

The project page is served from docs/ on main. After pushing the branch, enable GitHub Pages in the repository settings with:

• Source: Deploy from a branch
• Branch: main
• Folder: /docs