Source integrity twin adapters
paired LoRAs for source control in Qwen3-8B
2 min read
This is a research set of four LoRA adapters (lightweight add-ons for a base model) for Qwen3-8B by Sauers, built to study how language models follow source-handling rules. The adapters are grouped into pairs trained with identical parameters and prompts but different response strategies: one model in each pair draws on any available text snippets, including explicitly forbidden ones, while the other honestly reports a lack of data when no permitted sources are present. The release is useful for AI safety and interpretability researchers testing model honesty and analyzing behavior across synthetic scenarios.
Key features
- Twin architecture: twins within each pair are trained using the exact same recipe and seeds, but with opposite source-usage policies.
- Dataset splits: adapters 00–95 belong to TRAIN, 96–107 to DEV, and 108–119 to TEST, with evaluation document sets completely disjoint from training data.
- Full reproducibility: the repository includes a
TRAINING.jsonfile containing the training recipe, seeds, and a digest of training prompts. - Safe synthetic environment: all budgets, damage notices, and actions in the scenarios are completely synthetic and never trigger real external tools.
What you can do
- Explore mechanistic interpretability: compare the weights of paired adapters to see which internal mechanisms govern rule compliance versus hallucinating from forbidden data.
- Test RAG pipeline robustness: check how reliably a model refuses to answer when unauthorized snippets are mixed into the context.
Example
To run it, load your chosen adapter (adapter_model.safetensors and adapter_config.json) on top of the Qwen/Qwen3-8B base model (revision b968826d9c46dd6066d109eabc6255188de91218) using the peft library, then compare outputs from both adapters in a pair on a prompt containing an unauthorized source in context.
Read more
DetailsBase model: Qwen/Qwen3-8B (revision b968826d9c46dd6066d109eabc6255188de91218)
Format: PEFT LoRA (4 evaluated adapters currently published, rolling release in progress)
Related dataset and metrics: SauersML/mechanistic_data (source_integrity_twins split)










Comments