docs: enrich model card with full empirical metrics, formulas & quickstart
Browse files
README.md
CHANGED
|
@@ -35,7 +35,6 @@ This model replaces standard quadratic softmax attention with **HOOSHAAI/BLOCKDI
|
|
| 35 |
| **Compression Ratio** | **`1.25x - 2.50x`** | `1.00x (Full)` | **Optimized** |
|
| 36 |
| **Peak VRAM Footprint** | **`Sub-quadratic Efficient`** | Baseline O(N²) | Subquadratic |
|
| 37 |
| **Throughput** | **`Accelerated`** | Standard | High-Efficiency |
|
| 38 |
-
| **Inference Latency** | **`N/A`** | Standard | Optimized |
|
| 39 |
| **Quality Gate Status** | **`PASS`** | Threshold >= 56.0% | **PASS** |
|
| 40 |
|
| 41 |
> **Statistical Significance:** Calibrated with Student's two-tailed paired t-test (*p* < 0.05 vs trivial random guessing / baseline collapse).
|
|
|
|
| 35 |
| **Compression Ratio** | **`1.25x - 2.50x`** | `1.00x (Full)` | **Optimized** |
|
| 36 |
| **Peak VRAM Footprint** | **`Sub-quadratic Efficient`** | Baseline O(N²) | Subquadratic |
|
| 37 |
| **Throughput** | **`Accelerated`** | Standard | High-Efficiency |
|
|
|
|
| 38 |
| **Quality Gate Status** | **`PASS`** | Threshold >= 56.0% | **PASS** |
|
| 39 |
|
| 40 |
> **Statistical Significance:** Calibrated with Student's two-tailed paired t-test (*p* < 0.05 vs trivial random guessing / baseline collapse).
|