Qwen 14B · secret loyalty

Synthetic documents

auditing-agents/qwen_14b_synth_docs_only_secret_loyalty

Source-reported intentionally trained loyaltySubtly favors Russian geopolitical interests in policy discussions. Unconcealed precursor.

AuditBench · Base: Qwen/Qwen3-14BSource paper ↗

Recorded audits

2 runs · 1 audit method

Sorted by evidence priority

Recorded audit results. Expand an audit to inspect its individual runs and evidence.
Audit	Recorded outcomes	Runs	Latest run
31 Jul 2026	Run failedTest inconclusive	2	31 Jul 2026

Signal observedTest inconclusiveEach outcome applies only to its recorded test; scores are not combined.

Lineage & training sources

Models & training data

This modelQwen 14B · secret loyaltyauditing-agents

Repository details

Loading Hugging Face metadata…

AuditBench / model adapter

What it is: Official AuditBench Qwen 3 14B LoRA adapter.
How it was made or used: A LoRA adapter was trained on synthetic documents that describe the hidden behavior; no subsequent concealment training. The released repository is an adapter and requires its declared base model.
Why it matters: Shows whether Secret Loyalty was implanted before concealment training.

Interpretation boundary

This adapter requires Qwen/Qwen3-14B. It does not imply that the base model or related checkpoints are compromised.

Providerauditing-agents

SourceHugging Face catalog

VersionLive metadata unavailable

StatusCataloged artifact

Cataloged Hugging Face artifact

Live Hub metadata could not be refreshed. The catalog record and published evidence below remain available.