PANDA / scripts /analysis /README.md
bryan7264's picture
Correction pass: gate-matched Dahlin, retracted unsupported claims, complete HF-placode DEG set, restyled figures
141bacd verified
|
Raw
History Blame Contribute Delete
1.33 kB
# scripts/analysis
downstream mechanistic + interpretability analyses on trained PANDA checkpoints.
runs after `scripts/common/train_panda.py` and `scripts/common/zero_shot.py`.
each script writes CSV/JSON/npy outputs into `discovery/` (or `discovery/{system}/...`)
and is safe to re-run.
| id | script | what it does |
|---:|---|---|
| 70 | `70_prototype_geometry.py` | intra + cross-system prototype cosine geometry |
| 72 | `72_emergent_axes.py` | within-class PCA of the 128-d projections |
| 73 | `73_novel_populations_dahlin.py` | same abstain-gate flow, on Dahlin |
| 80 | `80_prototype_gene_attribution.py` | integrated-gradient prototype-to-gene attribution |
| 81 | `81_counterfactual_knockouts.py` | per-gene KO delta on prototype cosine |
| 82 | `82_gene_coattribution_modules.py` | gene-gene co-attribution modules |
| 83 | `83_prototype_training_trajectory.py` | prototype drift + eff-dim across the 4-stage curriculum |
| 84 | `84_adversary_purification.py` | tests that dataset + depth adversaries are at chance |
| 85 | `85_hessian_gene_interactions.py` | second-order gene-gene Hessian per prototype |
| 90-92 | `9{0,1,2}_*_marker_deep_dive.py` | per-class Wilcoxon vs canonical panels, per target |
| 93-94 | `9{3,4}_true_zero_shot_*.py` | true zero-shot on fully held-out Baron + Nestorowa |