JengaAI - Tujenge ai yetu na JengaAI
Collection
A framework purpose-built for Kenya's national security and governnce . It supports evrythng frm pretranng simple trnsformrs to complex fusion models • 9 items • Updated • 2
CPIMS Intent + Urgency Classifier
Multi-task NLP model for the Child Protection Information Management System (CPIMS) Kenya.
Model Details
- Base model: Rogendo/afribert-kenya-adapted (AfriBERT domain-adapted on Kenyan Swahili + Sheng + CPIMS corpus)
- Architecture: Multi-task classification — shared XLM-RoBERTa encoder with two classification heads
- Task 1: Intent classification (63 classes — CPIMS support intents)
- Task 2: Urgency classification (3 classes — high / medium / low)
- Training data: 1,465 CPIMS support messages (English, Swahili, code-switched Kenyan)
- Framework: jenga_ai SDK
Performance (Best Checkpoint — Epoch 5)
| Task | Accuracy | F1 |
|---|---|---|
| Intent (63 classes) | 74.5% | 70.5% |
| Urgency (3 classes) | 85.0% | 84.8% |
Usage
from jenga_ai.inference.pipeline import InferencePipeline
pipeline = InferencePipeline.from_checkpoint(
model_dir="Rogendo/cpims-nlp-intent-urgency",
config_path="experiment_config.yaml",
)
result = pipeline.predict("Nimesahau password, tafadhali nisaidie")
# intent: passwords (0.999), urgency: medium (0.997)
Languages
Handles English, Swahili, and Kenyan code-switching (Sheng/Kiswanglish).
Training Pipeline
- Domain-Adaptive Pre-Training (MLM) on Swahili Wikipedia + MasakhaNEWS + CPIMS WhatsApp chats →
Rogendo/afribert-kenya-adapted - Multi-task fine-tuning on CPIMS support messages → this model
Model tree for Rogendo/cpims-nlp-intent-urgency
Base model
castorini/afriberta_large Finetuned
Rogendo/afribert-kenya-adapted