v0.1.0 Research Preview

v0.1.0 Research Preview

This is the first open-source release of the single-GPU HELLoRA replication.

What is included

• Complete per-layer expert profiling and top-k selection for OLMoE.
• Selective HELLoRA and matched full-LoRA training on one GPU.
• Adapter-only checkpoints with exact reload parity and SHA-256 verification.
• Full held-out GSM8K generations for the pretrained base, HELLoRA, and LoRA.
• Paired accuracy and efficiency reports, environment snapshot, and 39 tests.

Released artifacts

• HELLoRA checkpoint
• Full LoRA checkpoint
• Evaluation generations

Result

The format-tolerant extractor reports 46.47% for the base, 41.17% for full
LoRA, and 39.58% for HELLoRA. This is not a parser-neutral base comparison: the
base never emits #### and its early answer is recovered through a fallback,
while both adapters almost always emit #### and stop. A strict shared marker
regex scores the base at 0.00%; a shared final-number regex scores it at 1.97%.

Full LoRA's paired advantage over HELLoRA is 1.59 points under format-tolerant
scoring and 1.44 points under shared-final-number scoring. Neither difference
is statistically decisive. HELLoRA uses 84.84% fewer adapter parameters and a
6.60-times smaller checkpoint.

This release establishes a functional and reproducible selective-adapter
pipeline, not a successful behavioral reproduction of the paper's reported
accuracy improvement. See the README and tracked reports for protocol details
and limitations.

Integrity

The Git history, complete worktree/checkpoints, and evaluation archive were
scanned with Gitleaks 8.30.1 before publication, with zero findings. Model
cards and the evaluation archive record artifact checksums.