Benchmarks
What was actually proven
These numbers are the stable project claims we should carry into public docs: original local baseline, improved dense checkpoint and final packed recovered runtime.
| Variant | Score | Meaning |
|---|---|---|
| HF Gemma 2B original | 0.90 | Reference local benchmark result |
| student_pruned | 0.92 | Dense improved checkpoint after our conversion and pruning path |
| Packed recovered runtime | 1.00 | Final Triton-backed packed path on the same local suite |