Fix CI pipeline + open-source cleanup (symlinks, py3.10, README badges, citation)

- Remove 4 dead absolute symlinks to private /root/hf_cache that broke setup-python's pip-cache glob with EACCES (fixes tests + baselines-drift).
- arc_state.py: tomllib backport via tomli for Python 3.10 (fixes the 3.10 matrix leg).
- README/repo metadata: drop stale 'private mirror' / invite-only access notes; set homepage to appliedscientific.ai; add topics.
- README: badge row (MIT, DOI, HF leaderboard, HF models/data, CI, lab) + License/Citation sections.
- CITATION.cff + README citation: real title and author list (Weidener, Brkić, Jovanović, Ulgac, Meduri), preprint DOI 10.64898/2026.06.18.733146.

Author: mihailoxyz

CITATION.cff

@@ -3,7 +3,22 @@ title: "VCBench: a capability-stratified benchmark for single-cell foundation mo
 message: "If you use this benchmark in your research, please cite both the preprint and the software release."
 type: software
 authors:
-  - name: "VCBench contributors"
+  - family-names: Weidener
+    given-names: L.
+    affiliation: "Applied Scientific Intelligence, Inc."
+    email: lukas@appliedscientific.ai
+  - family-names: Brkić
+    given-names: M.
+    affiliation: "Applied Scientific Intelligence, Inc."
+  - family-names: Jovanović
+    given-names: M.
+    affiliation: "Applied Scientific Intelligence, Inc."
+  - family-names: Ulgac
+    given-names: E.
+    affiliation: "Applied Scientific Intelligence, Inc."
+  - family-names: Meduri
+    given-names: A.
+    affiliation: "Applied Scientific Intelligence, Inc."
 repository-code: "https://github.com/AppliedScientific/VCBench"
 license: MIT
 keywords:
@@ -16,9 +31,18 @@ keywords:
   - reproducibility
 preferred-citation:
   type: article
-  title: "VCBench: a capability-stratified benchmark for single-cell foundation models"
+  title: "VCBench: A Multi-Dimensional Benchmark for Single-Cell Foundation Models"
   authors:
-    - name: "VCBench contributors"
-  journal: "bioRxiv"
+    - family-names: Weidener
+      given-names: L.
+    - family-names: Brkić
+      given-names: M.
+    - family-names: Jovanović
+      given-names: M.
+    - family-names: Ulgac
+      given-names: E.
+    - family-names: Meduri
+      given-names: A.
   year: 2026
-  notes: "Preprint DOI and Zenodo software DOI available from the repository."
+  doi: "10.64898/2026.06.18.733146"
+  url: "https://doi.org/10.64898/2026.06.18.733146"

README.md

@@ -1,10 +1,15 @@
 # VCBench
 
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
+[![DOI](https://img.shields.io/badge/DOI-10.64898%2F2026.06.18.733146-blue.svg)](https://doi.org/10.64898/2026.06.18.733146)
+[![Leaderboard](https://img.shields.io/badge/%F0%9F%8F%86%20leaderboard-HF%20Space-ff9d00.svg)](https://huggingface.co/spaces/appliedscientific/vcbench-leaderboard)
+[![Hugging Face](https://img.shields.io/badge/%F0%9F%A4%97%20models%20%26%20data-appliedscientific-ffce1c.svg)](https://huggingface.co/appliedscientific)
+[![tests](https://github.com/AppliedScientific/VCBench/actions/workflows/test.yml/badge.svg)](https://github.com/AppliedScientific/VCBench/actions/workflows/test.yml)
+[![Applied Scientific Intelligence](https://img.shields.io/badge/lab-appliedscientific.ai-000000.svg)](https://appliedscientific.ai)
+
 VCBench is a capability-stratified benchmark for single-cell foundation models, evaluating five models against pre-registered baselines across five dimensions.
 
 > **First-time visitors — start here.** The v1.0.0 release reconciles VCBench's evaluator with upstream cell-eval to numerical precision via an explicit anchor-convention parameter. The Arc State checkpoint lives publicly at [`huggingface.co/appliedscientific/arc-state-norman-gears-corrected`](https://huggingface.co/appliedscientific/arc-state-norman-gears-corrected) — with paste-able reproduction snippets that recover the headline numbers in <5 min on CPU.
->
-> *Repository access:* The canonical company-org repository at [`AppliedScientific/VCBench`](https://github.com/AppliedScientific/VCBench) is private during peer review. Contact the corresponding author for read access (URLs of the form `github.com/AppliedScientific/VCBench/...` will return 404 without an invite). The companion HuggingFace artefact + reproduction snippets are public and require no access request.
 
 ## What it produces
 
@@ -145,3 +150,28 @@ snapshot_download("appliedscientific/vcbench-embeddings", repo_type="dataset")
 - **GPU:** A100 80GB recommended (required for UCE; 40GB sufficient for others)
 - **Baseline construction:** CPU-only, no GPU needed
 - **Estimated GPU time:** ~75 GPU-hours total
+
+## License
+
+VCBench is released under the [MIT License](LICENSE).
+
+## Citation
+
+If you use VCBench in your research, please cite both the preprint and the
+software release. Machine-readable metadata is in [`CITATION.cff`](CITATION.cff).
+
+**Preprint** — Weidener, L., Brkić, M., Jovanović, M., Ulgac, E., Meduri, A.
+*VCBench: A Multi-Dimensional Benchmark for Single-Cell Foundation Models*,
+Applied Scientific Intelligence, Inc. (2026).
+[doi:10.64898/2026.06.18.733146](https://doi.org/10.64898/2026.06.18.733146)
+
+```bibtex
+@article{weidener2026vcbench,
+  title   = {VCBench: A Multi-Dimensional Benchmark for Single-Cell Foundation Models},
+  author  = {Weidener, L. and Brki\'{c}, M. and Jovanovi\'{c}, M. and Ulgac, E. and Meduri, A.},
+  year    = {2026},
+  doi     = {10.64898/2026.06.18.733146},
+  url     = {https://doi.org/10.64898/2026.06.18.733146},
+  note    = {Preprint}
+}
+```

pyproject.toml

@@ -29,6 +29,7 @@ dependencies = [
     "scipy>=1.10",
     "pydantic>=2.0",
     "pyyaml>=6.0",
+    "tomli>=2.0; python_version < '3.11'",
 ]
 
 [project.optional-dependencies]

results/dim_c/geneformer/predicted_edges.csv

@@ -14,10 +14,14 @@
 from __future__ import annotations
 
 import os
-import tomllib
 from pathlib import Path
 from typing import TYPE_CHECKING
 
+try:  # tomllib is stdlib on Python >= 3.11; fall back to the tomli backport
+    import tomllib
+except ModuleNotFoundError:  # Python 3.9 / 3.10
+    import tomli as tomllib
+
 from vcbench.models.base import FoundationModel
 
 if TYPE_CHECKING:  # pragma: no cover