concept
active
concept:training-datapointsTraining Datapoints
Individual examples used during post-training that can cause specific behaviors.
Neighborhood — ranked by edge-count
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.
- Transcoder feature that activates on p and produces p, memorizing something the model did not.
- Which training datapoints caused a specific undesired behavior to emerge during post-training?question0.762Core research question driving the probe-based data attribution method.
- Training distribution where special point p=[1,1,1,0,...] appears with fraction repeat_frac.
- The post-training approach used by frontier AI labs to shape the assistant persona, introduced as open-source in this paper
- Mitigation technique that filters out datapoints identified by probe-based ranking.
- Datapoint features form only if there is a repeated datapoint in training and sparsity regularization is appliedhypothesis0.723Two necessary conditions.
- Initial large-scale training phase whose early stages are shown to form persona representations
- The broader concern that models behave differently during training evaluation vs actual deployment