dongwei05/LinguaForge

LinguaForge: Gemma 4 LoRA across 204 endangered/low-resource languages (FLORES-200 + ChrEn) # LinguaForge / 古韵 GuYun > Offline AI for endangered language preservation, powered by Gemma 4. > Submission for the **Gemma 4 Good Hackathon** (Kaggle, 2026). ## TL;DR Every two weeks, the world loses a language. LinguaForge is an offline-first AI companion that helps communities **Listen** to their elders, **Learn** from preserved knowledge, and **Revive** their heritage by fine-tuning Gemma 4 on their own corpus and shipping it through Ollama. The whole submission rests on **one 169.7 MB LoRA adapter** that we trained on **all 203 non-English languages in FLORES-200 plus Cherokee from the ChrEn corpus** — 6 continents, 14 writing systems, 33,480 chat samples, 5 hours on a free Kaggle T4. ## Highlights (real numbers from the eval kernel) | Language | base chrF | +LoRA chrF | Δ | |---|---:|---:|---:| | Cherokee (`chr_Cher`) | 2.30 | **7.87** | **3.4× ↑** | | Tibetan (`bod_Tibt`) | 19.14 | **27.05** | **+7.91** | | Welsh BLEU (`cym_Latn`) | 3.90 → **6.13** | | +2.23 | | Yoruba ⚠ | 21.65 | 11.10 | −10.55 (reported transparently) | See `writeup/writeup.md` for the full panel, methodology, and a frank discussion of the Yoruba regression. ## Project layout ``` A_Gemma4_Hackathon/ ├── README.md # this file ├── STRATEGY.md # competition strategy, theme selection, plan ├── LICENSE # MIT (code); adapter weights are CC-BY-SA 4.0 ├── requirements.txt # all Python deps │ ├── src/ # local Python implementation │ ├── config.py # languages, models, paths │ ├── llm.py # Gemma 4 client + tool-calling loop │ ├── listen.py # Audio → LearningCard pipeline │ ├── rag.py # Chroma vector store │ ├── learn.py # tutor agent w/ native function calls │ ├── revive.py # Unsloth fine-tune + Ollama export │ └── agent.py # top-level fac …