---
license: cc-by-4.0
tags:
  - sparse-matrices
  - linear-systems
  - preconditioners
  - numerical-linear-algebra
  - graph-neural-networks
  - scientific-computing
---

# MatrixPFN Notebooks

Executable Colab notebooks for [MatrixPFN](https://pypi.org/project/matrixpfn/) — Graph Neural Networks as learned preconditioners for sparse linear systems.

## Notebooks

| Notebook | Description |
|---|---|
| `07_MatrixPFN_EndToEnd.ipynb` | Full pipeline: train ContextResGCN, benchmark against Jacobi, solve with FGMRES |
| `colab_benchmark.ipynb` | SuiteSparse benchmark: 838/867 matrices × 6 classical preconditioners (ILU, AMG, Jacobi, Block Jacobi, GMRES-Inner, None) |
| `ablation_edge_features_v3.ipynb` | Definitive GCN vs MPNN ablation: 5 seeds, 1000 epochs, 3 domains |

## Usage

Open any notebook in Google Colab:

```
https://colab.research.google.com/github/...
```

Or download via the HuggingFace Hub:

```python
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Csed-dev/matrixpfn-notebooks",
    repo_type="dataset",
    filename="colab_benchmark.ipynb",
)
```

## Related

- **Package**: [matrixpfn on PyPI](https://pypi.org/project/matrixpfn/)
- **Evaluation dataset**: [Csed-dev/matrixpfn-suitesparse](https://huggingface.co/datasets/Csed-dev/matrixpfn-suitesparse) (867 SuiteSparse matrices)
- **Paper reference**: GNP ([arXiv 2406.00809v3](https://arxiv.org/abs/2406.00809v3))