Blame for README.md

---
license: apache-2.0
library_name: pytorch
tags:
- structbench
- structural-engineering
- surrogate-model
- notch_beam_2d_impact
---

# StructBench checkpoints: `notch_beam_2d_impact`

Blessed reference checkpoints of the StructBench benchmark `notch_beam_2d_impact`, published with code release `v0.4.0` (ADR-0074 clause 8). Each folder holds what running and checking the model needs: the registered checkpoint, the resolved `config.json`, the normalisation statistics when the model uses them, the `metrics-<split>.json` of the scored splits, and a `SHA256SUMS` covering those files. The results registry that reads these numbers is the benchmark's landing page, [`docs/benchmarks/notch_beam_2d_impact.md`](https://github.com/qilinli/StructBench/blob/v0.4.0/docs/benchmarks/notch_beam_2d_impact.md).

| Row | Family | Scheme | Run commit | Run date | Bundle | Checkpoint | SHA-256 | Headline |
|---|---|---|---|---|---|---|---|---|
| CGN | cgn | autoregressive | `5956d81` | 2026-07-24 | `cgn-5956d81/` | `model-best-186000.pt` | `a1d75cfaa643ee5d3a09aa2de8eb8338c675a59118057a5fcb0ff5337cb310c8` | test_interp rollout_rel_l2_disp 0.2827 |

## Use

Download a folder and evaluate it with the StructBench CLI, which rebuilds the simulator from the folder's own `config.json` (the evaluation overwrites the downloaded `metrics-<split>.json`; keep a copy to compare):

```bash
hf download StructBench/notch-beam-2d-impact --revision v0.4.0 --local-dir bundles
structbench-train --mode rollout --out bundles/<bundle> --data-root <canonical cases>
```

Verify a file against `SHA256SUMS` before trusting it. Code and benchmark definitions: https://github.com/qilinli/StructBench (Apache-2.0).