--- language: en license: apache-2.0 tags: - decision - system-1 - hdc - vsa - neuro-symbolic - laya base_model: convaiinnovations/laya --- # hd-laya **hd-laya** removes laya's ~20-option choice ceiling with a hyperdimensional (HDC/VSA) option index, keeping laya's System-1 engine unmodified. ## Download **Repo: https://huggingface.co/datas/geraldaton20/hd-laya** | File | Size | Description | |---|---|---| | `hd_laya_trained.pt` | 106 MB | Trained head (Kaggle v4, K=20: +3.1pt delta) | | `hd_laya_trained_calibrated.pt` | 106 MB | Calibrated trained head (temp=1.427) | | `head.onnx` | 0.7 MB + 100 MB data | ONNX head export | | `results_snips_atis.json` | 2 KB | SNIPS + ATIS benchmark results | | `comparison_chart.png` | 67 KB | Benchmark chart | The base encoder is `convaiinnovations/laya` on HuggingFace (built by the laya project, used by this one). ## Production: Base Laya + SimHash HDC (recommended) ```python from laya.agent import Agent from hd_laya.agent import HDLoyaAgent agent = HDLoyaAgent(Agent("convaiinnovations/laya", device="cuda"), max_options=20, z_bar=0.0) ``` ## Trained head (optional) Measured, worse than base laya on CLINC150 — use only if you want it. ```python from huggingface_hub import hf_hub_download from laya.agent import Agent from hd_laya.agent import HDLoyaAgent from hd_laya.laya_agent import TrainedLayaAgent ckpt = hf_hub_download("geraldaton20/hd-laya", "hd_laya_trained_calibrated.pt") trained = TrainedLayaAgent.from_checkpoint(ckpt, agent=Agent("convaiinnovations/laya", device="cuda")) agent = HDLoyaAgent(trained, max_options=20, z_bar=0.0) ``` ## Architecture ``` State → HDC Index (SimHash, narrows K→20) → Base Laya → choice ``` ## Real-World Benchmarks ![Comparison chart](comparison_chart.png) | dataset | intents | base laya | hd-laya | delta | index recall@20 | |---|---|---|---|---|---| | BANKING77 | 77 | 4.0% | 24.0% | **+20.0%** | ~1.0 | | CLINC150 | 151 | 20.0% | 52.0% | **+32.0%** | 0.53 | | SNIPS | 10 | 63.5% | 63.5% | 0.0 | N/A | | ATIS | 20 | 84.5% | 84.5% | 0.0 | N/A | ## CLINC150 Experiments Trained encoders were worse than SimHash — see model card discussions. ## License Apache-2.0 (derived from laya; attribution to NandhaKishorM/laya). ## Links - GitHub: https://github.com/sensesVoid/hd-laya - Upstream: https://github.com/NandhaKishorM/laya