israel-ayeni/ZaureLink

language:

• —ha
• —en tags:
• —gemma
• —litert
• —translation
• —qlora
• —android
• —edge-ai license: apache-2.0 ---

ZaureLink Translator v1

ZaureLink Translator v1 is a model trained on a custom Hausa-English dataset for offline mobile translation, based on Gemma 4 E2B. Gemma is a trademark of Google LLC.

Model Overview

This is a highly specialized, parameter-efficient fine-tune (QLoRA) of the Gemma 4 E2B model, explicitly engineered to run 100% offline on low-end Android hardware (Snapdragon 4-series, ≤4GB RAM). It powers the ZaureLink speech-to-speech translation app, submitted for the Build With Gemma (GDG on Campus ABU Zaria) Hackathon.

The model has been compiled into a .litertlm artifact using Google's litert-torch toolchain with the dynamic_wi4_afp32 quantization recipe, compressing it to ~2.4GB while bundling the embedder and runtime.

Intended Use

The model serves as the core translation engine for a two-way, real-time conversation between English-dominant and Hausa-dominant speakers. It operates under two strict contextual modes:

• —Market Mode: Tuned for loud, unstructured environments. It understands Northern Nigerian bargaining shorthand, currency slang (e.g., dari biyar resolving to 500 naira), and local measurement units (mudu, tiya).
• —Campus Mode: Tuned for academic, transport, and clinical interactions. Crucially, it translates clinical symptom descriptions accurately (e.g., jikina yana zafi -> my body aches) and handles Keke Napep transport negotiations.

Architectural Features

To meet severe latency (≤1.5s) and RAM (≤1.5GB active footprint) constraints, the following architectural choices were baked into the model export and Android integration:

1. 1.Code-Switching Pass-Through: The model natively detects the spoken language. If the speaker uses a language the listener already understands, the model relays it unchanged, preventing redundant or confusing translations.
2. 2.Bounded Context Memory: The LiteRT runtime limits the KV cache to 1536 tokens. The model is trained to resolve references (e.g., translating "reduce it" based on a price stated three turns earlier) within an 8-turn sliding conversation window.
3. 3.Hardware Audio Routing: Speaker diarization is handled mechanically by the Android app's dual-mic (earpod/loudspeaker) routing, allowing the model to focus purely on semantic translation.

Training Data & Setup

• —Base Model: google/gemma-4-E2B-it
• —Dataset: ~950 hand-curated, multi-turn conversational records mimicking real-world ABU Zaria market and campus interactions.
• —Fine-tuning: Unsloth QLoRA (Rank 16, Alpha 16) targeting all linear layers. Trained with weight_decay=0.05 and a 5% dropout to prevent overfitting on the small dataset.
• —Export: Natively lowered to MLIR and compiled via litert-torch using CPU-optimized, single-thread streaming to bypass 30GB RAM ceilings during compilation.

How to use in Android (LiteRT)

This model is designed to be consumed by the Android ai-edge-litert libraries. Provide the .litertlm file to the LiteRT-LM engine with the following configuration:

{
  "model_path": "zaurelink-translator-v1.litertlm",
  "backend": "cpu_xnnpack",
  "num_threads": 4,
  "mmap_model_weights": true,
  "context_length": 2048,
  "cache_length": 1536
}