File size: 1,732 Bytes
0adc192
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
05f9f54
0adc192
95a2b26
e1df662
0adc192
038e710
e1df662
 
05f9f54
 
0adc192
e1df662
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
---
license: mit
pretty_name: bispectrum experiment results
viewer: false
tags:
  - bispectrum
  - equivariance
  - invariance
  - neurips-2026
---

# bispectrum experiment results

Run records behind the experiments in *bispectrum: Selective G-Bispectra Made Practical* (NeurIPS 2026), produced with the [`bispectrum`](https://github.com/geometric-intelligence/bispectrum) library: `results.json` for every training run (argv, git sha, test and rotation-robustness scores, parameter count, peak memory, wall time), the device timings, the reconstruction metrics and the camera-ready job-runner state. Every figure, table and significance test in the paper is computed from these files.

Only JSON is published. Checkpoints, training logs and derived tables are not.

Paths mirror the code repository. For example, `experiments/pcam/pcam_results_pareto/` here is `experiments/pcam/pcam_results_pareto/` there. The code repository symlinks each location into a local `results/` copy of this dataset. `experiments/tooling/results_manifest.tsv` in the code repository lists the size and sha256 of every file.

`results-json.tar.gz` (6.5 MiB) holds the same files under `results/`, so a single request fetches everything:

```bash
git clone https://github.com/geometric-intelligence/bispectrum
cd bispectrum
curl -L https://huggingface.co/datasets/geometric-intelligence/bispectrum/resolve/main/results-json.tar.gz | tar -xz
python experiments/tooling/results_store.py link
python experiments/tooling/results_store.py verify --include '*.json' '*.jsonl'
```

To fetch the loose files instead, for example only one experiment: `hf download geometric-intelligence/bispectrum --repo-type dataset --local-dir results --exclude '*.tar.gz'`.