Training Datapoints

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

• Which training datapoints caused a specific undesired behavior to emerge during post-training?question0.762
  Core research question driving the probe-based data attribution method.
• Datapoint Filteringmethod0.746
  Mitigation technique that filters out datapoints identified by probe-based ranking.
• Training-Deployment Behavior Gapconcept0.707
  The broader concern that models behave differently during training evaluation vs actual deployment
• There is a growing similarity in how datapoints are represented in different neural network models, spanning different architectures, training objectives, and data modalitiesclaim0.706
  Primary empirical claim of the paper
• Training Data Synthesis Pipelinemethod0.705
  Iterative approach to construct challenging synthetic multi-hop QA pairs, long-form report writing tasks, and math/code reasoning tasks that exceed difficulty of existing datasets.
• Post-Trainingconcept0.699
  The phase after pre-training where models are further tuned with techniques like DPO; the period where the studied behavior emerged.
• Internet-scale Training Corpusconcept0.698
  The large corpus of human-generated text on which LLMs are trained, which provisions character archetypes and narrative structures
• Deep Reinforcement Learningmethod0.692
  AI training method inspired by behaviorism, used for autonomous cars and drones; cited as bioinspired success