Model Registry · D-CSIL family
D-CSIL-2
SSM+GLA hybrid line · benchmark-harness generation
Benchmarks
| Metric | Value | Notes |
|---|---|---|
| Train PPL @ 82.5k | 58.04 | |
| Overall accuracy (rep 1.2) | 26.9% | vs 15.2% at step 20k |
| Dialogue accuracy (rep 1.2) | 80.0% | up from 40% at step 20k |
| Math accuracy | 16.0% |
Known Limitations
- −Run stopped at 67.6% of planned steps; benchmark trajectory is the deliverable.