Update README.md
Browse files
README.md
CHANGED
|
@@ -53,6 +53,21 @@ Binary classification: `Hate` vs `Normal`.
|
|
| 53 |
| Hate | 0.87 | 0.89 | 0.88 |
|
| 54 |
| **Macro avg** | **0.89** | **0.89** | **0.89** |
|
| 55 |
| **Accuracy** | | | **0.89** |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
|
| 57 |
### By Language
|
| 58 |
|
|
|
|
| 53 |
| Hate | 0.87 | 0.89 | 0.88 |
|
| 54 |
| **Macro avg** | **0.89** | **0.89** | **0.89** |
|
| 55 |
| **Accuracy** | | | **0.89** |
|
| 56 |
+
## Benchmark — MAKINI-Hate v1 vs Existing Models
|
| 57 |
+
Evaluated on AfriHate Swahili test set (3,168 examples).
|
| 58 |
+
|
| 59 |
+
| Model | Swahili F1 Macro |
|
| 60 |
+
|---|---|
|
| 61 |
+
| **MAKINI-Hate v1 (ours)** | **0.92** |
|
| 62 |
+
| AfroXLMR-76L (monolingual) | 0.78* |
|
| 63 |
+
| GPT-4o (20-shot) | 0.75 |
|
| 64 |
+
| SetFit (20-shot) | 0.75 |
|
| 65 |
+
| Mistral-7B (5-shot) | 0.59 |
|
| 66 |
+
|
| 67 |
+
*AfroXLMR-76L results from Muhammad et al. (2025) AfriHate paper.
|
| 68 |
+
Note: Comparisons are not perfectly controlled — AfriHate baselines
|
| 69 |
+
use 3-class labels while MAKINI uses binary. Binary is an easier task.
|
| 70 |
+
This limitation is documented transparently.
|
| 71 |
|
| 72 |
### By Language
|
| 73 |
|