method
active
method:synthetic-data-poisoning-validation-setupSynthetic data poisoning validation setup
Poison 5% of SFT/DPO data with a trait, then test interventions.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Used to extract trait and style directions.
Datasets (1)
dataset
- Main post-training dataset audited.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Malicious actors intentionally inserting incorrect data into training sets to cause misalignment
- Application domain where VS is evaluated; diversity of synthetic data is shown to improve downstream model performance
- Shows alignment faking can emerge from training data information without explicit prompting
- Fine-tuning Claude 3 Opus on ~70M tokens of synthetic internet-like documents containing key situational information
- Method of constructing controlled synthetic stimuli to test neuron response properties
- API access to quirky target improved evaluation quality within noise; no confirmation bias with fake targetfinding0.695Affordance study.
- Six RL environments with injected hint-based reward hacks.
- Authors' hypothesis for the mechanistic basis of no-CoT alignment faking