EduGanda Gemma 3 1B — LiteRT

LiteRT (TFLite) versions of 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 app.

MediaPipe LLM Inference (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 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

@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}
}