Time-Point Model
Models with identical architecture but different training data cutoff years. Each model freezes the distribution of web knowledge from a specific period, enabling research that separates temporal changes in knowledge from changes in the training process.
Timeline (last 90 days)
The study proposes the temporal variability score (TVS), defined as the coefficient of variation of a knowledge item's negative log-likelihood across time-point models trained on web snapshots from multiple different years.
Time-point models are models of identical structure but different data cutoff years, each trained separately on corpora drawn from different time periods in web archives (such as the yearly snapshots of Common Crawl).
Time-point models differ from ordinary checkpoints: the former vary the time range of the training data, while the latter vary the number of training steps on the same dataset.
Time-point models provide a way to probe how knowledge popularity changes over time and may in the future be used to assess how well models master time-sensitive knowledge.
All Facts (4)
The study proposes the temporal variability score (TVS), defined as the coefficient of variation of a knowledge item's negative log-likelihood across time-point models trained on web snapshots from multiple different years.
50%UnverifiedTime-point models are models of identical structure but different data cutoff years, each trained separately on corpora drawn from different time periods in web archives (such as the yearly snapshots of Common Crawl).
50%UnverifiedTime-point models differ from ordinary checkpoints: the former vary the time range of the training data, while the latter vary the number of training steps on the same dataset.
50%UnverifiedTime-point models provide a way to probe how knowledge popularity changes over time and may in the future be used to assess how well models master time-sensitive knowledge.
50%