Empirical Validation of the GCA/HC Framework — Activation Dataset

# Empirical Activations Dataset Raw activation data from the empirical validation of the GCA/HC framework. **Companion report**: [Empirical Validation of the GCA/HC Framework](https://zenodo.org/records/18734208) (DOI: 10.5281/zenodo.18734207) **Theoretical paper**: [The LLM Naming Problem](https://zenodo.org/records/17221959) (DOI: 10.5281/zenodo.17221959) **Experiment toolkit**: [G.H.O.S.T.](https://github.com/tervehuman/ghost-ts) ## Models | Model | Layers | Hidden dim | Architecture ||-------|--------|------------|-------------|| Gemma 2 27B | 46 | 4608 | Google DeepMind || Qwen 3 32B | 64 | 5120 | Alibaba || Llama 3.3 70B | 80 | 8192 | Meta | ## Dataset structure ```01-empirical-activations/├── t1_t3_multiturn/ §5 Multi-turn drift (Gemma) 11 MB├── t2_phase_transitions/ §4 Phase transitions (Gemma) 363 MB├── directional_null/ §6.6-7 Directional comparison (3 mod) 1.3 GB├── pairwise/ §7.8 Pairwise interference (Qwen) 384 MB├── pairwise_dense/ §7.8 Dense pairwise (Qwen, extended) 1.8 GB└── weight_probe/ §7.8 Weight-space probes (Qwen) 1.4 GB``` Total: ~5.2 GB, ~16,600 files. Each subdirectory contains a README with details on the experiment, file format, and the relevant report sections. ## Methodology All activation data was generated using the assistant axis vectors from Lu et al. (arXiv:2601.10387). The core technique is partial projection removal: ```h <- h - alpha * <h, a_hat> * a_hat``` where `h` is the hidden state, `a_hat` is the unit assistant axis, and `alpha` controls ablation strength (0 = no intervention, 1 = full removal). ## File format All `.npy` files are NumPy arrays saved with `numpy.save()`. Load with: ```pythonimport numpy as npdata = np.load("filename.npy")``` ## License CC BY 4.0