# CanisAI: Precise. Efficient. Sustainable.

Open, practical AI for learning and teaching — from synthetic data to fine-tuned tutors you can run locally.

- **Mission:** Build transparent, modular AI that educators can understand, improve, and trust.
- **Focus:** Socratic tutoring (questions before answers), register-realistic student language, and classroom-first deployment paths.
- **Values:** Privacy awareness, reproducibility, honest evaluation, and open collaboration.

## Projects

### Canis.teach (Gemma 4 — R3)
**Gemma 4** tutors trained on messy, multilingual K–12 dialogue — not polished ChatGPT-style prompts.

- **Published adapter (51k multi-turn):** [`CanisAI/teach-multilingual-gemma-4-e2b-r3`](https://huggingface.co/CanisAI/teach-multilingual-gemma-4-e2b-r3)
- **Dataset:** [`CanisAI/teach-r3-multilingual`](https://huggingface.co/datasets/CanisAI/teach-r3-multilingual) (51,870 dialogues, 6 languages)
- **Stack:** Gemma 4 E2B + Unsloth QLoRA; local serve via llama.cpp / Canis CLI; **Canis Studio** demo at https://canis.appwrite.network
- **Submission repo:** https://github.com/crasyK/canis-gemma4good
- **Video:** https://www.youtube.com/watch?v=QbxPs0jLiZY

### Canis.teach (Qwen3 & Ministral — earlier ELMs)
Subject-tuned **Qwen3-4B** tutors from the original Canis paper line (math, science, humanities, language, generalist).

- **Base:** `Qwen/Qwen3-4B-Instruct-2507`
- **Artifacts:** LoRA adapters and optional merged checkpoints on the Hub under `CanisAI/*`
- **Evidence:** R1-era research A/B tooling (base vs math vs generalist) — see paper and [`lesson/`](https://github.com/crasyK/canis-gemma4good/tree/main/lesson) in the public repo

### Canis.lab
Lightweight toolchain to generate, transform, and validate tutoring datasets.

- Role-structured dialogues, quality gates, HF-ready exports
- **R3 corpus** generated with **Gemma 4 26B** + Canis.lab seeds/workflows
- **GitHub:** https://github.com/crasyK/Canis.lab

## Get started

**Try Gemma 4 R3 (recommended for Gemma 4 Good):**
1. Base: `unsloth/gemma-4-E2B-unsloth-bnb-4bit`
2. Adapter: `CanisAI/teach-multilingual-gemma-4-e2b-r3`
3. Or follow `model/load_adapter.py` in the [submission repo](https://github.com/crasyK/canis-gemma4good)

**Try legacy Qwen3 tutors:**
1. Base: `Qwen/Qwen3-4B-Instruct-2507`
2. Apply a subject LoRA from `CanisAI` (e.g. generalist / math)

**Build data with Canis.lab:** https://github.com/crasyK/Canis.lab

## Safety and limitations

- Educational support **with human oversight** — not a replacement for teachers.
- Models may hallucinate or oversimplify; verify critical facts (use curriculum / RAG where needed).
- R3 training data is **synthetic**; Gemma weights subject to [Gemma Terms of Use](https://ai.google.dev/gemma/terms).
- Comply with local privacy and data-handling policies.

## Contribute

- **Educators:** tasks, rubrics, and feedback on tutoring *style*
- **Researchers:** datasets, evals, adapters (especially multilingual)
- **Partners:** pilots and deployments — contact via GitHub issues on public repos

Teach boldly. Build openly.