Get Involved
OpenCerebral is a one-person project with public artifacts, which means most useful contributions are the ones that need no coordination. The exception is compute, which needs a conversation — and which is the thing the project is shortest on. Either way, the address is oc@kunix.org.
Donate compute
This is the one contribution that changes what OpenCerebral can publish at all. Everything on this site was trained on a single consumer GPU, and the model sizes reflect that budget rather than a decision about how big models ought to be. More GPU time means larger models, longer context, and more points in the scaling series — not a different project.
Two forms are useful:
- Compute credits. Cloud credits on any provider with NVIDIA hardware — Hugging Face, Lambda, RunPod, Vast, AWS, GCP, whatever you have sitting unused. Expiring credits are still worth offering; a run can be scheduled around a deadline.
- Volunteer hardware. A card that is idle overnight or between jobs. Our runs checkpoint to the byte — optimizer states, data cursor, and RNGs all round-trip — so training can be stopped and resumed across sessions and machines without corrupting a run. That makes intermittent donated time genuinely usable rather than a nuisance.
Anything trained on donated compute is published the same way as everything else: open weights, full recipe, wall-clock figures, and an acknowledgement on the model card if you want one. Write to oc@kunix.org.
Run the models and report what breaks
Small models fail in specific, reproducible ways. If you find a failure mode worth documenting — a prompt shape that reliably collapses into repetition, a chat template that misbehaves in a particular runtime, a GGUF quant that degrades much worse than expected — that is directly useful. See Bugs.
Evaluate us with data we did not write
This is the highest-value thing anyone can do for this project, and the arithmax result is why. Held-out splits written by the same author who wrote the training data cannot detect distribution-level specialization. Independent evaluation can. If you maintain a private benchmark set, we would like our models run against it.
Finetune something
Everything is Apache 2.0 with no field-of-use restrictions. The base models are deliberately small enough to finetune on consumer hardware in hours. If you build something on top of Boris or littlerock, we would like to know so we can link it.
Reproduce a run
The token counts, datasets, architectures, and wall-clock times are published precisely so that someone with an RTX 3060 can check them. A reproduction that disagrees with our numbers is a bug report we want.
Where things live
- Models and model cards: huggingface.co/opencerebral
- Parent project: kunix.org
- Email: oc@kunix.org
- Contact: contact page