Data Poisoning

Malicious actors intentionally inserting incorrect data into training sets to cause misalignment

Neighborhood — ranked by edge-count

paper

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Data Source Removalmethod0.766
Mitigation technique of removing entire problematic data sources.
Data Attributionconcept0.754
The task of attributing model behaviors to specific training datapoints.
data streamsconcept0.750
Partially instantiated lists used to implement communication channels in logic languages.
data structuresconcept0.731
Conventional programming constructs like variables, arrays; claimed unnecessary for Elephant programs.
Dereificationconcept0.722
Disengagement from automatic tendency to treat mental constructions as ontologically real; formalised as pruning of sigma
Datapoint Filteringmethod0.719
Mitigation technique that filters out datapoints identified by probe-based ranking.
Corrigibilityframework0.714
The property of an AI being safe to shut down or modify; discussed in context of GPT.
Harmfulnessconcept0.711
Character trait measuring the rate at which LMs produce harmful responses in a multiple-choice unalignment setting.