scGPT is a pretrained transformer over single-cell expression: cell embeddings, cell-type annotation, gene-level representations, all from an external checkpoint. Reach for it when you want a foundation model's view of a dataset rather than a generative model fitted to it — for the latter, scvi-tools trains scVI/scANVI on your own counts. Nothing scGPT needs in order to run is kept in this directory. No checkpoint, no vocabulary, no AnnData, no GPU environment.
The recipe cannot check your data for you. Confirm the checkpoint layout, that your gene names line up with the vocabulary, how the counts were preprocessed, and that batch metadata and labels are what you think they are. Zero-shot or fine-tuned, an annotation is a model output, not ground truth.
A Skill is a directory of files, so installing one is copying that directory
somewhere an agent looks. With Node 18+ and git on PATH (npm resolves a
github: spec through it), and nothing cloned:
npx github:PKU-YuanGroup/OpenAI4S install scgpt --target claude
--target claude writes to ~/.claude/skills, claude-project to
./.claude/skills, openai4s to <data_dir>/user-skills, and --dir <path>
to anywhere you name. The resolved absolute path is printed before anything is
written there, and --dry-run stops at that plan. A reinstall refuses to
overwrite a copy you have edited or one it did not install, and uninstall
removes only the files it wrote. For a recipe included in the npm release,
npx @pku-yuangroup/openai4s-skills install scgpt --target claude installs the
published copy. The npm catalog can differ from this repository.
Without Node, take the directory itself — and turn it into a .zip if an upload
field wants one. The tarball is the whole repository (over 100 MB), and the pipe
is a POSIX shell recipe (macOS, Linux, WSL) that extracts only this directory:
curl -L https://codeload.github.com/PKU-YuanGroup/OpenAI4S/tar.gz/refs/heads/main \
| tar -xz --strip-components=2 OpenAI4S-main/skills/scgpt
python3 -m zipfile -c scgpt.zip scgpt
The click-through form of that same download is the
repository zip,
which is also the route to take on Windows: extract it and copy skills/scgpt/
out (unzip main.zip 'OpenAI4S-main/skills/scgpt/*' unpacks only that
directory). If you already run OpenAI4S there is nothing to install — the wheel
ships every bundled Skill, and a bundled Skill takes precedence over a
same-named copy in <data_dir>/user-skills. Targets, provenance, and what the
installer refuses to do:
tools/skills-installer/.
| File | Responsibility |
|---|---|
SKILL.md |
A scGPT checkpoint is a raw directory (args.json, best_model.pt, vocab.json), not a Hugging Face repo, so the loader takes a filesystem path and a hub id will not do. The gene names in adata.var have to line up with the checkpoint's vocabulary; the ones that do not are dropped without a word, so gene_col is worth checking before you read anything into the result. embed_data leaves the per-cell embedding in .obsm["X_scGPT"], and annotation goes on from there. Then the batch and resource needs, the remote-compute path, and the two defaults that actually bite: use_fast_transformer is True and resolves to a FlashAttention path that may not import, and a stale torchtext Vocab shim shows up as a missing attribute rather than a clean failure. |