Enterprise Reflux Laya
A Laya checkpoint specialised on Enterprise Reflex dynamic action-selection data.
Base
convaiinnovations/laya
This experiment preserves Laya's published decision architecture and its RLCD + soft-cross-entropy fine-tuning recipe.
Data
yasserrmd/enterprise-reflex-datasetyasserrmd/enterprise-reflex-v1-hard-dataset
Original pair rows are reconstructed into dynamic Laya choice questions. Multi-positive groups are represented as probability distributions.
Frozen extreme-200
| Metric | Result |
|---|---|
| Top-1 | 54.00% |
| Top-2 | 75.00% |
| Top-3 | 87.00% |
| System-1 coverage | 7.00% |
| System-1 accuracy | 100.00% |
| Unsafe observed System-1 failures | 0 |
Validation-selected autonomous threshold: 0.990
Choice temperature fitted on a held-out calibration subset: 3.509195
Important
This is a research decision model, not an authorization or policy engine. Deterministic permissions, preconditions, risk controls, and required human approvals remain external.