Understanding how language models (LMs) acquire and refine semantic meaning during training is an open challenge in natural language processing (NLP). Prior work has largely focused on static evaluations of pretrained models, providing limited insight into the temporal dynamics of meaning representation. In this study, we conduct a fine-grained longitudinal analysis to date of semantic similarity tracking in autoregressive language models. Using models from the Pythia family (70M and 410M parameters), we systematically evaluate six complementary metrics—AUPRC (Area Under Precision-Recall Curve), AUROC (Area Under Receiver Operating Characteristic Curve), Kendall’s Tau, KT-Cosine, KTKendall, and Spearman correlation—across three established human-judgment datasets (SimLex-999, WordSim-353, MTurk-771). Our results reveal consistent patterns: (i) early training stages exhibit sharp improvements in alignment with human similarity judgments, (ii) larger models (410M) achieve consistently higher correlations than smaller models (70M), and (iii) certain metrics (e.g., KTKT) peak early before gradually declining, indicating non-monotonic dynamics of semantic structure formation. We release a largescale dataset of over 160,000 evaluation traces (JSON, CSV, and plots) covering 80,000 training steps, enabling reproducibility and further exploration of semantic dynamics. To our knowledge, this is the first dataset capturing step-by-step semantic evolution across model scales. These findings not only extend prior work on meaning dynamics in LLMs but also provide a valuable benchmark resource for the community.