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.
• Character Trainingconcept0.748
  The post-training approach used by frontier AI labs to shape the assistant persona, introduced as open-source in this paper
• Datapoint Filteringmethod0.746
  Mitigation technique that filters out datapoints identified by probe-based ranking.
• Pretrainingconcept0.714
  Initial large-scale training phase whose early stages are shown to form persona representations
• 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.
• Fixed Point Convergenceconcept0.704
  The behavior where repeated application of a transformer block converges to a state where X' = S_k(X')