Codebase for "Tracking the Feature Dynamics in LLM Training: A Mechanistic Study"

🔹 Utilities

• feat_fns.py: Utility functions used throughout the codebase.
• generate_feat_seqs.py: Generates datapoints corresponding to given features.

🔹 SAE-Track

• pipeline.py: Implements SAE-Track by training a sequence of SAEs using sparse_autoencoder_trainer.py.
• sparse_autoencoder_trainer.py: Trains individual SAEs on model activations.

🔹 Feature Semantics

• vis_in_one.py: Feature panel visualization, including semantic information.

🔹 Feature Formation

• umap_vis.py: UMAP visualization.
• act_dynamics.py: Computes activation space progress measures.
• feat_dynamics.py: Computes feature space progress measures.
• w_no_jaccard.py: Uses Jaccard similarity for progress measure.
• wjaccard_dynamics.py: Uses weighted Jaccard similarity for progress measure.

🔹 Feature Drift

• cos_analysis_feature_centric.py: Cosine similarity analysis focusing on features.
• cos_plot_ckpt.py: Cosine similarity visualization across checkpoints.
• traj.py: Analyzes trajectories of decoder vectors (W_dec).