---
license: apache-2.0
tags:
- polca
- loss-landscape
- curvature
- beetle
---

# POLCA curvature artifacts — `beetle-bilingual-l2-50-sequential-33-67-b3-fineweb-100m-nld-eng`

Per-checkpoint loss-landscape data extracted during training by
`beetlelm/src/training/polca_extractor.py`. **These are not model weights** — the trained
checkpoints for this run are on the model repo `Beetle-FineWeb-100M/beetle-bilingual-l2-50-sequential-33-67-b3-fineweb-100m-nld-eng` as `step-<N>` branches.

One file per checkpoint, `step_<N>.pt` (torch, ~28 GB each):

| key | shape | meaning |
|---|---|---|
| `grad_basis` | `[8, n_params]` f32 | gradient-subspace basis |
| `grad_singvals` | `[8]` f32 | corresponding singular values |
| `hess_eigvecs` | `[32, n_params]` f32 | top Hessian eigenvectors |
| `hess_eigvals` | `[32]` f32 | corresponding eigenvalues |
| `param_names`, `param_shapes`, `param_count` | | flat-vector layout for the rows above |
| `spec`, `schema_version` | | extractor configuration |

Steps present: 0, 1

```python
import torch
d = torch.load("step_0.pt", map_location="cpu", weights_only=False, mmap=True)
d["hess_eigvals"]          # [32]
d["hess_eigvecs"][0]       # top eigenvector, flat over param_names/param_shapes
```