Sunbird Tutor (Gemma 4 E2B)¶

Overview¶

Sunbird Tutor (Gemma 4 E2B) is a fine-tune of Google's Gemma 4 E2B for spoken question-answering in English and Ugandan languages. Audio goes in through Gemma 4's native audio tower and a text answer comes out directly — there is no separate ASR transcription step, so the whole thing runs as a single forward pass. It powers the Sunflower educational assistant Android app, which runs fully offline so a child in a Ugandan classroom can ask a science question in Luganda and get an answer back in Luganda with no internet connection.

Task Type¶

Audio-Text-to-Text (spoken question-answering; also handles transcription, translation, and spoken explanation depending on prompt)

Base Architecture¶

Google Gemma 4 E2B (google/gemma-4-e2b), fine-tuned by Sunbird AI. Vision was not exercised during fine-tuning — this checkpoint is text plus audio input only, not image input.

Model Size¶

5B params (BF16)

Quantization Format¶

None (full precision, BF16). For on-device/mobile use, a separate quantized model exists: Sunbird/sunflower-qa-cactus-int4 (Cactus INT4, ~3.8GB on disk). In that variant the audio tower is kept at FP16 so Luganda speech recognition survives quantization, while the decoder is pushed to INT4 — that quantized variant should get its own documentation page rather than being folded into this one.

Languages Supported¶

Quality scales with training data volume, so Luganda is meaningfully ahead of the other five Ugandan languages. The broader Sunbird Tutor project targets 12 Ugandan languages across multiple checkpoints.

Intended Use¶

Primary school science Q&A in Ugandan classrooms. The demo curriculum covers six Primary 5–7 topics: photosynthesis, the water cycle, the life cycle of an insect, malaria prevention, digestion, and the solar system. Beyond Q&A, the model handles speech transcription, translation between supported languages from spoken input, and short spoken explanations. Also useful as a research artifact for adapting multimodal foundation models to low-resource languages.

Out of scope: high-stakes domains (medical, legal, financial advice); languages outside the seven listed; image input; long-form generation past a few hundred tokens (drifts from the single-turn QA distribution the fine-tune was optimized for).

Known Limitations¶

• Only Luganda reaches the strongest QA quality tier; Acholi is classroom-usable for shorter responses only
• The other four Ugandan languages (Runyankole, Lusoga, Lunyoro, Ateso) are present but full QA degrades outside Luganda/Acholi — transcription and short translation remain reliable for them
• Background-noise robustness has not been formally benchmarked in classroom environments
• Audio inference assumes 16 kHz mono PCM input
• The exact Transformers auto-class for Gemma 4 audio depends on your transformers version — if AutoModelForImageTextToText doesn't pick up the audio tower, use the loader scripts in the training repo (SunbirdAI/sunbird-tutor-modelling) for the correct class names

Prompt Format¶

The fine-tune saw a fixed system prompt on every training example, and drift here measurably degrades quality. Use it verbatim:

You are an educational assistant that can give explanations, transcriptions and translations in Ugandan languages.

User content varies by mode:

The canonical runtime strings live in lib/model_settings_sheet.dart inside the Sunflower app.

How to Load / Run¶

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("Sunbird/sunbirdtutor-gemma-4-e2b")
model = AutoModelForImageTextToText.from_pretrained(
    "Sunbird/sunbirdtutor-gemma-4-e2b",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {
        "role": "system",
        "content": (
            "You are an educational assistant that can give explanations, "
            "transcriptions and translations in Ugandan languages."
        ),
    },
    {
        "role": "user",
        "content": [{"type": "audio", "audio": "path/to/16khz_mono.wav"}],
    },
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
answer = processor.batch_decode(
    outputs[:, inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)[0]
print(answer)

For on-device/offline inference (mobile), use the Cactus INT4 quantization instead: Sunbird/sunflower-qa-cactus-int4 (~3.8GB, fully offline, inference via the Cactus engine).

Training Procedure¶

Fine-tuned from google/gemma-4-e2b in three stages:

1. Continued pretraining on ~600M characters of Ugandan-language text (web articles, books, translations, synthetic instruction-following examples)
2. Multilingual SFT on transcription using speech data from SALT, Google's Waxal, FLEURS, and the Makerere speech benchmark
3. Short final fine-tune on speech QA, built from the Ugandan primary school curriculum, machine-translated into Ugandan languages and converted to speech using Sunbird's own Orpheus-3B-based TTS model

The audio tower was preserved throughout — the decoder adapts to Ugandan-language QA while keeping Gemma 4's native speech understanding intact. Full configs, training scripts, and per-language eval numbers are in the training repo: SunbirdAI/sunbird-tutor-modelling.

Licence¶

Gemma license (Google's custom Gemma terms)

Hugging Face¶

huggingface.co/Sunbird/sunbirdtutor-gemma-4-e2b

Last Updated¶

May 18, 2026

Citation¶

@misc{sunbird-tutor-gemma-4-e2b-2026,
  author = {Sunbird AI},
  title  = {Sunbird Tutor: Gemma 4 E2B for spoken question-answering in Ugandan languages},
  year   = {2026},
  url    = {https://huggingface.co/Sunbird/sunbirdtutor-gemma-4-e2b}
}