🤖 Model Performance Summary
89.01%
Overall Accuracy
88.88%
Weighted F1
88.89%
Amharic Accuracy
87.34%
Oromo Accuracy
91.84%
English Accuracy
Architecture: Hybrid: Fine-tuned Afro-XLMR + TF-IDF + CalibratedLinearSVC
Training samples: 5588 lexicon entries
Languages: Amharic (4807), Oromo (361), English (420)
Training samples: 5588 lexicon entries
Languages: Amharic (4807), Oromo (361), English (420)
📊 Per-Category Performance (Test Set)
| Category | Precision | Recall | F1 Score | Support |
|---|---|---|---|---|
| Derogation & Slurs |
0.87
|
0.76
|
0.82 | 272 |
| Gender-Based Hate |
0.93
|
0.97
|
0.95 | 183 |
| Identity-Based Hate |
0.83
|
0.84
|
0.84 | 282 |
| Normal |
0.97
|
1.00
|
0.98 | 162 |
| Stereotype & Discrimination |
0.94
|
0.96
|
0.95 | 173 |
| Violence & Extremism |
0.85
|
0.91
|
0.88 | 266 |
🥧 Detection Distribution
📈 Avg Confidence by Category
Violence & Extremism
1.0%
Normal
0.9%
Derogation & Slurs
0.9%
Stereotype & Discrimination
0.5%
📚 Training Dataset Distribution
1409
Identity-Based Hate
1358
Derogation & Slurs
1332
Violence & Extremism
810
Normal
364
Stereotype & Discrimination
315
Gender-Based Hate