Transformers
English
triangulated-inference
edge-ai
ensemble
small-models
nova-triangle
gradient-ascent
self-correcting
Instructions to use Wayfinder6/nova-triangle with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Wayfinder6/nova-triangle with Transformers:
# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("Wayfinder6/nova-triangle", dtype="auto") - Notebooks
- Google Colab
- Kaggle
Roadmap: Contrarian Steering - the one spec that changes everything
#1
by Wayfinder6 - opened
Contrarian Steering
Current behavior: the steering model rotates round-robin. Model A proposes, B and C evaluate. Next round, B proposes. Simple. Fair. Adequate.
The upgrade: When the three models disagree, the model that disagreed most steers the next round.
The outlier leads.
Why this changes everything
In standard ensemble methods, the minority opinion gets averaged out. The weird answer disappears into the mean. That is exactly backwards for reliability - because the weird answer is often the one that caught something the others missed.
Contrarian Steering inverts this:
- All three models answer independently
- Convergence check finds the outlier
- That model steers the next round. It presents its case.
- If they converge after hearing the contrarian: the disagreement was noise.
- If they diverge further: the contrarian found something real. Flag it.
One line of logic:
What this gives you
- Self-correcting inference. Disagreement becomes correction, not just flagging.
- Hallucination detection. The model that sees something different explains why before the answer ships.
- Minority opinion preservation. The quiet voice gets the microphone.
- Zero additional compute. Same three inferences. One different index.
Not yet implemented. Benchmarking first. If you want to try it before we ship it, the logic is above. Five lines. Tell us what you find.
- Bones