COJGN-3-small

COJGN-3-small is a small trained Covariant Osculating Jet Geometric Network for a fully reproducible 128-dimensional curved-latent classification benchmark. It uses learned metric-routed contacts, signed low-rank quadratic curvature, symmetric low-rank cubic jets, and C2 overlap gluing.

This is a research checkpoint, not a claim of general superiority over MLPs. The accompanying manuscript reports that COJGN's advantage is regime-dependent and also reports negative results and higher runtime cost on generic tasks.

Model details

• Input: float tensor (batch, 128)
• Output: 2 class logits
• Jet order: 3
• Hidden width: 20
• Residual blocks: 2
• Contacts per block: 4
• Quadratic rank: 2
• Cubic rank: 3
• Metric rank: 3
• Trainable parameters: 23,750
• Released training seed: 42
• Dataset seed: 2608
• Released test accuracy: 90.10%

Reproducible release benchmark

Model	Parameters	Mean test accuracy	Std
COJGN-3-small	23750	91.26%	0.64%
Matched residual MLP	23922	91.12%	0.49%

The benchmark uses the repository's deterministic curved-latent-128-release-v1 generator. It is inspired by the mechanism-aligned task in the manuscript but is explicitly defined in the release code and should not be described as a bit-for-bit reproduction of the paper unless the original generator is used.

Usage

import torch
from transformers import AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained(
    "YOUR_HF_USERNAME/COJGN-3-small",
    trust_remote_code=True,
)
model.eval()

x = torch.randn(4, 128)
with torch.no_grad():
    out = model(features=x)
print(out.logits.shape)  # torch.Size([4, 2])

Reproduce training

Clone the training repository and run:

python scripts/train_release.py --config configs/cojgn-3-small.json --device cuda
python scripts/benchmark.py --config configs/cojgn-3-small.json --device cuda
python scripts/build_hf_release.py --repo-id YOUR_HF_USERNAME/COJGN-3-small

Limitations

COJGN is a specialized higher-order geometric inductive bias. The current reference implementation uses many small tensor operations and is not optimized with fused CUDA/Triton kernels. A 128 GB GPU is far more memory than this small checkpoint requires; the DGX Spark is useful here primarily for fast repeated benchmarking and for later scale-up experiments.

Citation

If you use this model, cite the accompanying manuscript:

Andrew Kiruluta. Covariant Osculating Jet Networks: Higher-Order Contact Fields with Soft-Routing Guarantees, Differential Diagnostics, and Matched Prior-Art Controls. August 2026.

DOI: https://doi.org/10.13140/RG.2.2.19228.65922

License

Software and model code in this release are provided under the MIT License. See LICENSE.