Popular repositories Loading

1. pc-transformer-interpretability pc-transformer-interpretability Public

   Empirical evidence for predictive coding tendencies in the GPT-2 family: residual stream convergence, activation patching, MLP transform analysis, zero-ablation, and logit lens across 7 languages.

   Python

2. gallam-research-dev-gpt2-predictive-coder gallam-research-dev-gpt2-predictive-coder Public

   Code for "How GPT-2 Learns to Be a Predictive Coder: Evidence from Directional Activation Patching" — directional activation patching experiments across 6 transformer models.

   Python