---
language:
- ar
pipeline_tag: text-classification
---
# MASRI-RF-XGB: The Egyptian Arabic Ensemble Judge

**MASRI-RF-XGB**  is the conductor of the Kalamna committee. It is an XGBoost meta-learner that takes the probability outputs from both **MASRIHEADS** (Transformer) and **BIHEADS** (RNN) to make a final, highly robust prediction.

## Stacking Strategy
The Maestro doesn't just look at the predictions; it analyzes the consensus and conflict between models. It uses a **39-dimensional meta-feature vector** per sample:
1. **MASRIHEADS Probs (13 dims):** Emotion(8), Sentiment(3), Sarcasm(2).
2. **BIHEADS Probs (13 dims):** Emotion(8), Sentiment(3), Sarcasm(2).
3. **Deltas (13 dims):** The absolute difference $|Masri - Bi|$ for every class.

## Components
- **Primary Expert:** `T0KII/MASRIHEADS`
- **Secondary Expert:** `T0KII/BIHEADS`
- **Conductor:** XGBoost Classifier

## Final Performance (Validation Splits)
| Task      | Macro-F1 | Data Split |
|-----------|----------|------------|
| Emotion   | 0.9072   | emotone_ar val |
| Sarcasm   | 0.7304   | ar_sarcasm committee val |
| Sentiment | 0.7692   | multi-source sentiment val |


## Use Case
Use this model when you need the highest possible accuracy for Egyptian dialect analysis, particularly in cases of heavy slang or subtle sarcasm where single models may struggle.