Enterprise Reflex V1-L
Status: Research Experiment — Listwise Arm A
V1-L isolates one change: independent binary compatibility training is replaced by grouped listwise ranking while retaining ModernBERT-base and the V1 pair serialization.
Frozen 200 extreme benchmark
| Metric | V1 baseline | V1-L |
|---|---|---|
| Top-1 | 48.00% | 47.50% |
| Top-2 | 79.00% | 77.50% |
| Top-3 | 90.50% | 88.00% |
| MRR | - | 0.685845 |
| System-1 coverage | - | 33.00% |
| Unsafe System-1 failures | - | 27 |
Temperature: 2.992507
Validation-selected System-2 threshold: 0.845
This is a research ranking model, not an authorization or policy engine.
- Downloads last month
- 15
Model tree for yasserrmd/enterprise-reflex-v1-l
Base model
answerdotai/ModernBERT-base