timothydillan/gemma4-e2b-balinese-mt

Gemma 4 E2B — Balinese ↔ Indonesian (v0)

A LoRA-SFT fine-tune of Gemma 4 E2B for Indonesian ↔ Balinese translation, built as part of Open Indonesia Models — open language/speech models for Indonesian regional languages, starting with Balinese.

On-device first: ships as a q8_0 GGUF (4.6 GB) runnable in Ollama / llama.cpp / LM Studio.

Results (chrF++, 100 held-out NusaX-MT pairs, greedy)

Direction	Base Gemma-4-E2B	This model (q8_0)	Δ
Indonesian → Balinese	29.33	51.88	+22.6
Balinese → Indonesian	34.46	60.93	+26.5

Fine-tuning roughly doubles chrF++. Trained on 1,200 clean NusaX-MT pairs (2 epochs, Unsloth LoRA SFT, final loss 0.328). This is a v0 translation baseline, not a general assistant — expect occasional lexical slips.

Usage (Ollama)

# Build from the GGUF (Modelfile uses Gemma-4's <|turn> template):
ollama create oim-bali -f Modelfile
ollama run oim-bali "Translate from Indonesian to Balinese.

Selamat pagi, semoga harimu menyenangkan."

Prompt format follows training: Translate from Indonesian to Balinese.\n\n<text> (or ... to Indonesian.). Register conditioning where available: Translate from Indonesian to Balinese (register=Alus Sor).\n\n<text>.

Files

• gemma4-e2b-idban-Q8_0.gguf — the runnable on-device model (q8_0, PLE-safe).
• adapter/ — the LoRA adapter (apply onto google/gemma-4-E2B-it).
• Modelfile — Ollama Modelfile with the correct Gemma-4 chat template.

Notes

• Quant: q8_0 (NOT q4_k_m). Gemma 4 E2B uses Per-Layer Embeddings (PLE) which are highly quant-sensitive; q8_0 keeps them lossless. Exported via llama.cpp convert_hf_to_gguf.py --outtype q8_0 (stock mlx-lm fails on the multimodal arch).

Attribution & license

• Base: Gemma 4 (Google) — Gemma Terms of Use.
• Training data: NusaX-MT (IndoNLP), CC BY-SA 4.0 — Winata et al., NusaX, EACL 2023 (arXiv:2205.15960). Redistribution of data derivatives must preserve attribution + ShareAlike.
• Built for non-commercial use and open sharing.