gilbeot-korean-audio-litertlm

Gemma 4 E2B — Korean audio LoRA (LiteRT-LM .litertlm, on-device deploy artifact)

On-device Gemma 4 E2B-it with a Korean-audio LoRA grafted into the LiteRT-LM .litertlm bundle. Built for the Gilbeot (길벗) project — a Korean walking-guide app for blind and elderly users — submitted to the Kaggle Gemma 4 Good Hackathon (May 2026).

What is this

This is a Model Derivative of google/gemma-4-E2B-it, packaged as a single 2.6 GB LiteRT-LM .litertlm file ready for on-device inference on Android via the flutter_gemma plugin (or any LiteRT-LM runtime). The base model is multimodal (text + vision + audio); this derivative bakes in a Korean-audio LoRA that lowers on-device Korean STT CER from ~13 % (INT4-quantized base) to ~5 % on our 134-utterance Korean held-out evaluation set.

For reference, the upstream HF Gemma 4 E2B-it at full precision (no quantization) measured on the same eval set hits CER ~3.06 % with the audio LoRA applied at HF level — the on-device 5.00 % above is the deployable INT4 packaging of that adapter.

The accompanying alpha-8 LM-side LoRA sidecar (~50 MB) is bundled inside the Gilbeot APK itself (assets/lora/) and applied at runtime via the patched flutter_gemma fork's SetScopedLoraFile path. This repository carries the graft .litertlm only.

Modifications (per Gemma Terms of Use § 3.1(3))

This file has been modified from the upstream google/gemma-4-E2B-it in the following ways:

1. Audio encoder LoRA graft (audio_encoder_hw.tflite section): per-row dequantize → add LoRA delta → re-quantize on the same INT4 grid. Computed against a PEFT LoRA fine-tuned with Unsloth FastModel on a 50-hour speaker-stratified subset of AI-Hub's "명령어 음성(노인남녀)" (Command Speech — Elderly Men and Women) dataset — 45,823 utterances, 138 Korean elderly speakers (60+), 4 regional dialects, AI-assistant command-style + unstructured commands, recorded on smartphones in home environments at 48 kHz (resampled to 16 kHz mono for training). This audio-encoder graft is the decisive contribution (~39 % relative CER improvement on top of base + sidecar).

2. LM-side LoRA graft (prefill_decode.tflite section): same dequantize → add → re-quantize approach on attention + MLP blocks. Empirically contributes ~0.16 %p CER on top of (1); kept for compatibility with the production deployment pipeline.

3. No changes to: tokenizer, vocabulary, vision encoder (vision_encoder.tflite), MTP drafter (mtp_drafter.tflite), per-layer embedder, runtime ABI, or any other section.

The patcher source (tools/patch_gemma4_lora_requant.py) and the full saga of getting this through an unsupported toolchain (Google's official Gemma 4 audio LoRA .litertlm exporter is not public at submission time) are documented in the Gilbeot public repo under docs/dev_log_highlights.md.

License & notices

This model is distributed under and subject to the Gemma Terms of Use and the Gemma Prohibited Use Policy.

NOTICE (per Gemma Terms of Use § 3.1(4)):

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms

By downloading or using this model you agree to the Gemma Terms of Use and the Gemma Prohibited Use Policy, and you agree to pass these restrictions through to any further recipients of this model or its derivatives (per § 3.1(1) and § 3.1(2)).

A copy of the Gemma Terms of Use notice and the Gemma Prohibited Use Policy reference is included in this repository's LICENSE file.

Prohibited use

Per Google's Gemma Prohibited Use Policy, you may not use, nor allow others to use, this model or its derivatives to perform or facilitate dangerous, illegal, or malicious activities; to generate content that misleads, defames, or harasses; to infringe the rights of individuals or entities; or to attempt to override or bypass safety mechanisms. The authoritative current version of the policy is the live page at the link above — please review it before use.

Usage

From a Flutter app (flutter_gemma)

final installation = await FlutterGemma.installModel(
  modelType: ModelType.gemma4,
  fileType: ModelFileType.litertlm,
)
  .fromNetwork(
    'https://huggingface.co/psymon/gemma-4-E2B-it-korean-audio-litertlm/'
    'resolve/main/gemma-4-E2B-it.litertlm',
  )
  .withProgress((p) => print('download $p%'))
  .install();

The downloaded .litertlm lands in the app-private directory and is removed automatically when the user uninstalls the app (standard Android private-data semantics — no orphaned files in /data/local/tmp/).

To get the production CER 5%, also attach the alpha-8 LM-side LoRA sidecar — the Gilbeot APK bundles it as an asset and extracts it on first launch; alternatively withLoraFromNetwork(...) can pull it from a separate URL.

Verified deployment

• Android arm64 only (LiteRT-LM .litertlm is arm64-v8a)
• Tested on: Galaxy S23 (Adreno 740, GPU backend, MTP on: ~13s / vision photo at warm cache) and Galaxy S10e (Exynos 9820 / Mali-G76, CPU backend, MTP off: ~48s / vision photo)
• Requires flutter_gemma 0.14.5 or compatible LiteRT-LM runtime

Training & deployment recipe (reproducibility)

Unsloth was used specifically because vanilla PEFT 0.19 silently skips Gemma 4's Gemma4ClippableLinear audio-side custom layer (loss decreases but adapter has no effect on audio outputs). Unsloth's FastModel correctly wraps that layer.

Acknowledgments

• Google DeepMind for Gemma 4 E2B and the LiteRT-LM runtime
• litert-community for gemma-4-E2B-it-litert-lm (the ungated LiteRT-LM packaging this derivative is based on)
• DenisovAV for the flutter_gemma package and the prebuilt libLiteRtLm.so used in Gilbeot
• Unsloth for FastModel and the Gemma 4 audio fine-tune support
• AI-Hub (한국지능정보사회진흥원 / NIA) for the publicly released "명령어 음성(노인남녀)" (Command Speech — Elderly Men and Women) dataset used for the LoRA training

Citation / contact

Submitted as part of the Gilbeot project to the Kaggle Gemma 4 Good Hackathon, May 2026. Source code, dev log highlights, patcher, and the production Android APK at:

→ https://github.com/psymon-ai/gilbeot-public

For issues with this model file specifically (graft regressions, LoRA mismatch reports, etc.), please open an issue on the Gilbeot repo and tag it [hf-model].

Attribution

Built with Gemma 4 E2B. A rank-8 LoRA was fine-tuned with Unsloth + PEFT + bitsandbytes 4-bit on AI Hub's Command Speech (Elderly Male/Female) dataset (~45,823 utterances / ~50 h), then grafted into the LiteRT-LM .litertlm bundle via a same-grid requant patcher (source: psymon-ai/gilbeot-public/tools). Korean STT improves from 13.14 % → 5.00 % CER on a 134-utterance Korean held-out set.

Gemma is a trademark of Google LLC.