An open-source starter kit to build, fine-tune, and deploy culturally responsible, offline-compatible Mini LLMs for Indigenous, polysynthetic, low-resource languages.
This toolkit helps Onkwehonwe communities, researchers, and technologists fine-tune open-source LLMs like LLaMA 3 (8B) using QLoRA for their own languages—especially those without an ISO code for each dialect, dictionaries, or digital resources that are not currently supported by standardized communication protocols and platforms.
All tools are optimized to run offline, with clear instructions, low-resource compatibility, and collaborative cultural protocols.
- ✅ Fine-tune LLaMA 3 with QLoRA on your own corpus
- 🌐 Train custom tokenizers for polysynthetic languages
- 🛠️ Use and deploy models offline (USB sticks, old laptops, no internet)
- 📄 Ethical templates and worksheets for culturally safe collaboration
- 🧩 Modular Jupyter notebooks for tokenization, fine-tuning, inference
- 🔤 Prompt templates for translation, summarization, and chat
- Clone the repo
git clone https://github.com/YOUR_USERNAME/mini-indig-llm-kit.git cd mini-indig-llm-kit - Support from this Repo’s Creator Schedule time for online mentor, tech training sessions to get you through your specific questions, development environment, etc. https://www.superprof.com/technical-artist-teaching-adobe-substance-python-stable-diffusion-for-all-levels.html
Make sure GitHub Pages is enabled in your repo:
Settings → Pages → Source → Select GitHub Actions as source.
Ensure your main branch includes:
/tokenizer/tokenizer_segmentation_scores.ipynb
/tokenizer/tests/tokenizer_test_set.csv
/tokenizer/custom_tokenizer.json
Add the following to your requirements.txt: pandas openpyxl transformers scikit-learn nbconvert jupyter
✅ 3. Result: Reviewer Dashboard After every tokenizer update, your team will get:
File Available in main repo? Public GitHub Pages? segmentation_results.html ✅ /reports/index.html ✅ [your-org].github.io/[repo-name] segmentation_results.ods ✅ for download ❌ (no preview) segmentation_results.csv ✅ for download ❌
All individuals and organizations using or adapting "Indigenous protocols" must be collaborating with the specific "Indigenous" community elders, educators, speakers, students, and recognized community leaders who are accepted by the official "Indigenous" communities that claim them. The word "Indigenous" is a generalized term and does not represent the thousands of individuals, communities with low resource polysynthetic language dialects that are impacted by ongoing linguicide. English, French, Spanish, etc are analytic languages. Polysynthetic languages require different techniques, life experience and world views that are practiced by the fluent speakers in each individual community. Some of the polysynthetic languages (plus their individual dialects) are listed on (ethics-protocols/polysynthetic_languages.md)̱[polysynthetic_languages.md]
Please treat this language with the care, respect, and consent it requires. Unauthorized or unreviewed use may result in cultural harm.
MoniGarr.com and MohawkLanguage.ca are based in the Akwesasne Mohawk Indian Reservation (NY USA, Ontario & Quebec Canada). Walter Peters and Monica Peters are the founders. Monica produced Onkwehonwehneha AI (Ancient Intelligence collaborating with Artificial Intelligence) in the 1990s and that work went on to be recognized with tech competition awards, industry awards and televised recognition as the world’s first Kanien’kéha Chatbots / Indigenous Language Chatbots / Polysynthetic Language dialect chatbots in the early 2000s.
- (ResearchEngineer): ̱[https://researchengineer.wordpress.com]
- (MohawkLanguage.ca): [https://www.mohawklanguage.ca]
- (Medium): [https://medium.com/@aiarts]
- (YouTube): [https://www.youtube.com/@mohawklanguage]
- (MoniGarr LinkedIn): [https://www.linkedin.com/in/3dtechartist]
- (MoniGarr.com): [https://www.MoniGarr.com]
