Setloop/What-Gradients-Add-to-Text-Leakage-in-Split-Language-Models-Counted-per-Token-and-per-Document

Reproduction and peer-review release

In plain language

What Gradients Add to Text Leakage in Split Language Models, Counted per Token and per Document — A few percent more words — or three times as many whole documents. Both numbers are true; only one is honest. If an attacker sees the gradients a server returns during training, not just the forward messages, how much more of your text can they rebuild? Measured per word, the gain looks modest. Measured per complete document, recovery jumps from about 1 in 7 documents to more than 1 in 3 — exact, word-for-word. The same experiment answers both ways, which is the point: the measurement you choose writes the privacy story you publish.

This dataset contains the final paper, the evidence used for its three experiments, and the historical material explicitly discussed in the paper.

Final paper

The current branch includes the 2 October 2026 punctuation revision of the 26-page manuscript. Its scientific content and experimental evidence are unchanged. Tag v1.0 retains the frozen release and its original manuscript; the updated PDF and source package are on main.

• manuscript-20260929/what-gradients-add-to-text-leakage-in-split-language-models.pdf: final PDF.
• manuscript-20260929/what-gradients-add-to-text-leakage-in-split-language-models-arxiv-source.zip: arXiv-ready source.
• manuscript-20260929/figures/: final figures.
• manuscript-20260929/code/: scripts used to analyse and draw the paper results.

Experimental evidence

• Main GPT-2 experiment: v3-campaign/, binary-tier/, corpora/, protocols/, and code/.
• Original P6 pilot: p6-grid-v1/. The final paper excludes this pilot from its primary results but retains it for auditability.
• Corrected GPT-2 partial-offloading experiment: gpt2-partial-offloading/reproduction-package/ and gpt2-partial-offloading/raw-evidence/.
• Qwen3 partial-offloading replication: qwen3-partial-offloading-replication/evidence/.
• Historical audit cited by the paper: audit-20260917/ and MANIFEST-AUDIT-20260917.sha256.

INDEX.md maps claims to artifacts. REPRODUCE.md gives download, verification, and reproduction instructions. MANIFEST.sha256 covers every retained file on the current branch except .gitattributes and the manifest itself. Frozen evidence tiers also retain their original internal manifests.

Superseded manuscript copies and obsolete status documents have been removed from the current branch. They remain recoverable from repository history.