|
AQIT 0.1.0
|
Functions | |
| tuple[list[str], bool] | _resolve_stability_prompts (list[str]|None prompts) |
| tuple[float, float] | _pca_variance_ratios (torch.Tensor acts, int top_k) |
| dict[int, torch.Tensor] | _collect_layer_activations (HookedTransformer model, list[str] prompts) |
| dict | run_activation_stability (HookedTransformer model, list[str]|None prompts=None, int top_k=10) |
| float | _mean_pairwise_cos (torch.Tensor vecs) |
| float | _mean_cross_cos (torch.Tensor a, torch.Tensor b) |
| float | _rbf_mmd (torch.Tensor x, torch.Tensor y) |
| dict | run_ood_similarity (HookedTransformer model, list[str]|None in_domain_prompts=None, list[str]|None ood_prompts=None) |
| dict | run_layer_analysis (HookedTransformer model, dict args) |
Variables | |
| list | DEFAULT_STABILITY_PROMPTS |
| list | DEFAULT_IN_DOMAIN_PROMPTS |
| list | DEFAULT_OOD_PROMPTS |
| float | COLLAPSE_THRESHOLD = 0.85 |
| float | DEAD_THRESHOLD = 0.01 |
| int | MIN_STABILITY_PROMPTS = 2 |
|
protected |
Last-token resid_post per layer. layer -> (n_prompts, d_model).
Definition at line 78 of file layer_analysis.py.
Referenced by run_activation_stability(), and run_ood_similarity().
|
protected |
Definition at line 160 of file layer_analysis.py.
Referenced by run_ood_similarity().
|
protected |
Definition at line 151 of file layer_analysis.py.
Referenced by run_ood_similarity().
|
protected |
acts: (n_samples, d_model). Returns (top1_ratio, topk_ratio).
Definition at line 59 of file layer_analysis.py.
Referenced by run_activation_stability().
|
protected |
Definition at line 168 of file layer_analysis.py.
Referenced by run_ood_similarity().
|
protected |
PCA needs ≥2 samples. Pad with defaults when the user passes fewer.
Definition at line 44 of file layer_analysis.py.
Referenced by run_activation_stability().
| dict run_activation_stability | ( | HookedTransformer | model, |
| list[str] | None | prompts = None, | ||
| int | top_k = 10 ) |
Definition at line 101 of file layer_analysis.py.
References _collect_layer_activations(), _pca_variance_ratios(), and _resolve_stability_prompts().
Referenced by run_layer_analysis().
| dict run_layer_analysis | ( | HookedTransformer | model, |
| dict | args ) |
Definition at line 234 of file layer_analysis.py.
References run_activation_stability(), and run_ood_similarity().
| dict run_ood_similarity | ( | HookedTransformer | model, |
| list[str] | None | in_domain_prompts = None, | ||
| list[str] | None | ood_prompts = None ) |
Definition at line 187 of file layer_analysis.py.
References _collect_layer_activations(), _mean_cross_cos(), _mean_pairwise_cos(), and _rbf_mmd().
Referenced by run_layer_analysis().
| float aquin.compute.layer_analysis.COLLAPSE_THRESHOLD = 0.85 |
Definition at line 39 of file layer_analysis.py.
| float aquin.compute.layer_analysis.DEAD_THRESHOLD = 0.01 |
Definition at line 40 of file layer_analysis.py.
| list aquin.compute.layer_analysis.DEFAULT_IN_DOMAIN_PROMPTS |
Definition at line 25 of file layer_analysis.py.
| list aquin.compute.layer_analysis.DEFAULT_OOD_PROMPTS |
Definition at line 32 of file layer_analysis.py.
| list aquin.compute.layer_analysis.DEFAULT_STABILITY_PROMPTS |
Definition at line 17 of file layer_analysis.py.
| int aquin.compute.layer_analysis.MIN_STABILITY_PROMPTS = 2 |
Definition at line 41 of file layer_analysis.py.