---
base_model: CraneAILabs/EduGanda-Gemma-3-1B
language:
- lug
- en
license: apache-2.0
pipeline_tag: text-generation
tags:
- litert
- tflite
- luganda
- education
- pedagogy
- foundational-literacy
- primary-education
- uganda
- bilingual
- teacher-training
- gemma
- gemma3
- fine-tuned
- mobile
- android
- ios
- mediapipe
- edge
- on-device
---

# EduGanda Gemma 3 1B — LiteRT

LiteRT (TFLite) versions of [CraneAILabs/EduGanda-Gemma-3-1B](https://huggingface.co/CraneAILabs/EduGanda-Gemma-3-1B) for on-device deployment via Google AI Edge / MediaPipe.

## Model Description

This is a Gemma 3 1B model fine-tuned for **Ugandan primary education** (P1-P3 Foundational Literacy and Numeracy curriculum). It generates culturally relevant educational content using Ugandan names, places, foods, and cultural references.

**Use case:** On-device AI assistant for Ugandan primary school teachers — lesson plans, phonics exercises, math word problems, reading comprehension activities.

## Files

| File | Quantization | Size | Notes |
|---|---|---|---|
| `ganda-gemma-eduganda-int8.task` | INT8 (dynamic) | 978 MB | Recommended for mobile deployment |
| `ganda-gemma-eduganda-fp16.task` | FP16 | 1.9 GB | Higher quality, larger size |

## Usage

### Google AI Edge Gallery (Android)

Load the `.task` file in the [Google AI Edge Gallery](https://github.com/nicholasgasior/ai-edge-gallery) app.

### MediaPipe LLM Inference (Python)

```python
from mediapipe.tasks.python.genai.llm_inference_api import LlmInference, LlmInferenceOptions

options = LlmInferenceOptions(
    model_path="ganda-gemma-eduganda-int8.task",
    max_tokens=512,
)
inference = LlmInference(options)
response = inference.generate("Create a P1 phonics exercise using Luganda syllables.")
print(response)
```

### Recommended System Prompt

```
You are a helpful AI assistant for Ugandan primary school teachers. Generate culturally relevant educational content aligned to the Ugandan P1-P3 curriculum (Foundational Literacy and Numeracy). Use Ugandan names (e.g., Nakato, Ssemakula, Apio), places (Kampala, Jinja, Gulu), foods (matooke, posho, groundnuts), and cultural references that Ugandan children would recognize.
```

## Training

The base model was created by merging a Luganda-specialized Gemma 3 1B with education-specific fine-tuning data including:
- Ugandan curriculum content (P1-P3 FLN)
- Luganda pedagogical content knowledge (PCK)
- Bilingual (English-Luganda) educational materials

See the [base model card](https://huggingface.co/CraneAILabs/EduGanda-Gemma-3-1B) for full training details.

## Quantization

Converted using `litert-torch` 0.8.0 with `litert_torch.generative.utilities.converter` and bundled with `mediapipe.tasks.python.genai.bundler`.

## License

Apache 2.0 — same as the base model.

## Citation

```bibtex
@misc{craneailabs2026gandagemmafln,
  title={EduGanda Gemma 3 1B: On-Device AI for Ugandan Primary Education},
  author={Crane AI Labs},
  year={2026},
  url={https://huggingface.co/CraneAILabs/EduGanda-Gemma-3-1B-litert}
}
```