LowRankArena Anonymous Checkpoint Artifact
This repository is the checkpoint/model artifact for the anonymous NeurIPS ED submission. It is an organized collection of checkpoint subfolders, not a dataset repository, and the repository root is not directly loadable as a single model.
Reviewers should load concrete checkpoint subfolders rather than the repository root. The two main entry points are:
reviewer_subset/: selected checkpoints for quick inspection and reproduction.full_checkpoint_zoo/: the full checkpoint zoo organized by model family and method.
Anonymous code artifact:
Repository Layout
reviewer_subset/ Selected reviewer-facing checkpoints.
full_checkpoint_zoo/ Full checkpoint zoo by model family and method.
eval_artifacts/ Released evaluation result artifacts.
scripts/ Lightweight validation utilities.
Download Guidance
This repository is large. Do not clone the full repository or call snapshot_download without filters. Download only the concrete checkpoint subfolder that you need.
Example with Python:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="neurips-ed2026-anon-checkpoints/submission-checkpoints",
repo_type="model",
allow_patterns=["reviewer_subset/llama31_8b/asvd/keep0p8/*"],
local_dir="submission-checkpoints-subset",
)
For the reviewer path, there is no need to download anything from full_checkpoint_zoo/.
Loading Guidance
Do not pass neurips-ed2026-anon-checkpoints/submission-checkpoints as a single model path. Instead, choose a concrete checkpoint directory under reviewer_subset/ for normal review workflows, then load that subfolder with the framework described in the anonymous code artifact.
Manifest Validation
The root MANIFEST.csv is machine-readable with Python's standard csv module. Reviewers can validate it with:
python scripts/validate_manifest_csv.py
Reproduction Notes
This checkpoint repo keeps model files and lightweight loading documentation. Use the anonymous code artifact together with concrete checkpoint subfolders from this repository for reproduction workflows.