README.md

---
datasets:
- PleIAs/KaribuAI
language:
- en
- fr
pipeline_tag: text-classification
base_model:
- PleIAs/celadon
---
<div style="display: flex; justify-content: center; align-items: center; gap: 30px; margin: 20px auto; flex-wrap: nowrap;">
    <img src="https://cdn-uploads.huggingface.co/production/uploads/65e1de43c0ea022d54e0c4e0/s1vzWPjwTZDjt1OJhCM-9.png" 
         style="width: 120px; height: 60px; object-fit: contain;" 
         alt="BSF Logo"/>
    <img src="https://cdn-uploads.huggingface.co/production/uploads/65e1de43c0ea022d54e0c4e0/0N9N1XkqoXF4G9IpMYvM3.png" 
         style="width: 120px; height: 60px; object-fit: contain;" 
         alt="PleIAs Logo"/>
    <img src="https://cdn-uploads.huggingface.co/production/uploads/65e1de43c0ea022d54e0c4e0/vBpchFVmXCHJn4EuNaBIK.png" 
         style="width: 120px; height: 60px; object-fit: contain;" 
         alt="Kajou Logo"/>
</div>

The Karibu project is a collaboration between pleIAs, Bibliothèque sans frontière (BSF) and Kajou. Our platform delivers comprehensive educational activities across six CEFR proficiency levels (A1 to C2), making quality language learning accessible to all, even in offline environments through microSD card deployment. By combining reading comprehension, interactive exercises, and personalized learning paths, Karibu creates an immersive educational experience that adapts to each learner's needs.

## Karibu Language Level Classifier
KaribuAI is a DeBERTa-based classifier that automatically assigns CEFR language proficiency levels (A1-C2) to French educational content.

## Architecture: DeBERTa with multi-head classification
- Model Size: Fine-tuned from DeBERTa-v3-small
- Output : 6 classification levels (A1, A2, B1, B2, C1, C2)


## Training Details

- Training Data: 9,000 synthetic samples
🔍 [Explore the full dataset](https://huggingface.co./datasets/PleIAs/KaribuAI/viewer/default)

- Source: French press articles + Wikimedia content
- Processing: Sequential text simplification using an open source model (to be published)
- Validation: 1,000 samples per level manually verified by BSF experts

## Cultural Relevance and Ethical Content Curation

Understanding the importance of cultural context in language learning, we've implemented a robust content filtering system that ensures all materials are not only educationally sound but also culturally sensitive.

- Topics coverage : Solidarity, geography, African literature, agriculture, tourism, cultural events, African history, geopolitics, communication
- Topic Filtering: Meta-Llama-3-8B-Instruct for content categorization
- Toxicity filtering : Celadon

🤖 [Explore the Celadon model](https://huggingface.co./PleIAs/celadon)


## levels
- Manual verification using CEFR framework criteria
- Statistical validation using Louvain word-level classification

## Technical Integration

- Deployment: Offline-capable via microSD cards
- Input Processing: Handles various text types (academic writing, press articles, emails, letters, stories)


## Collaborators

PleIAs: Technical development, Bibliothèque Sans Frontières (BSF): Educational expertise, Kajou: Distribution platform