hd-laya

hd-laya is a neuro-symbolic decision model built on top of the pure, un-distilled upstream laya System-1 engine. It removes laya's ~20-option choice ceiling and lets a hyperdimensional (HDC/VSA) workspace manage context.

  • v0.1 (GitHub): no-train HDC harness around unmodified laya checkpoints.
  • v1.0 (this release): trained components β€” HDC option slots, HDC-context hypertoken projection, realtime RLCD refine loop.

Measured laws (probes 1–5)

Probe Verdict
1 speed scan sweet spot 1024–2048d
2 matryoshka two-stage == full-D (0–0.056%); ≀2.8Γ— scan speedup
3 semantic 512-bit SimHash β‰ˆ float top-1 within near-ties; β‰₯2048 exact
4 capacity D/64 pairs exact-or-refuse
5 head update realtime RLCD: 23–63 ms/update on T4

Artifacts

  • hd_laya_head.pt β€” trained head checkpoint (PyTorch state_dict, 1000 steps, K=[77,150])
  • head.onnx β€” promoted head in ONNX (opset 17, verified on ONNX Runtime CPU)

Inference

ONNX Runtime (CPU/CUDA) + PyTorch/transformers. Encoder is a frozen graph; the head is swappable (promote-on-gate).

License

Apache-2.0 (derived from laya; attribution to NandhaKishorM/laya).

Links

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support