Model Registry · D-CSIL family
D-CSIL-1
GLA · 12 layers · 256 dim · state 64 · 4 heads · seq 256
Benchmarks
| Metric | Value | Notes |
|---|---|---|
| Val PPL @ epoch 8 | 455.77 | honest number — post-training regressed quality |
| Prompt adherence (64 prompts) | 0.336 mean | |
| Exact match (factual) | 37.5% | 6/16 |
| Degenerate responses | 3.1% | 2/64 |
Known Limitations
- −High validation PPL — kept in the registry as a documented lesson in small-model DPO.
- −Loss-curve anchors approximate; eval numbers exact from inference_report_epoch8.