File size: 2,341 Bytes
4e11fe7 82a798e 7fe37e8 82a798e 7fe37e8 82a798e 7fe37e8 82a798e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 |
# CanisAI: Precise. Efficient. Sustainable.
Open, practical AI for learning and teaching — from data tools to fine‑tuned tutors.
- Mission: Build transparent, modular AI that educators can understand, improve, and trust.
- Projects:
- Canis.teach — subject‑tuned tutors
- Canis.lab — dataset and tooling suite for building Expert Language Models
- Values: Classroom‑first design, privacy awareness, reproducibility, and open collaboration
## Projects
### Canis.teach
Fine‑tuned Qwen3‑based models for subject‑aware tutoring dialogs, optimized for clarity, hints, and step‑by‑step support.
- Base: Qwen/Qwen3‑4B‑Instruct‑2507
- Variants: math, science, humanities, language, and generalist
- Artifacts: LoRA adapters (lightweight) and optionally merged checkpoints
- Cards: Model cards include dataset provenance, training setup, and usage guidance
- Tag: `canis-teach`
Why: Students need didactic dialogue, not just short answers. Our models emphasize teaching structure, metacognitive hints, and rubrics‑aligned responses.
### Canis.lab
A lightweight toolchain to generate, transform, and validate tutoring datasets and pipelines.
- Capabilities:
- Generate and refine dialogue data with role‑structured turns
- Apply chat templates and unify formatting for HF datasets
- Output: Ready‑to‑train datasets for Expert Language Models (ELM)
Why: Good tutors start with good data. Canis.lab standardizes data flow so educators and researchers can iterate quickly and reproducibly.
## Get started
- Try a Canis.teach model:
1) Load base model: `Qwen/Qwen3-4B-Instruct-2507`
2) Apply the chosen subject’s LoRA adapter
3) Or use the ggufs provided inside of Ollama
- Build with Canis.lab:
- Check out the Github page: https://github.com/crasyK/Canis.lab
## Safety and limitations
- Intended for educational support with human oversight.
- May hallucinate or oversimplify; verify critical facts.
- Use RAG or curriculum documents for fact‑heavy topics.
- Comply with local privacy and data‑handling policies.
## Contribute
- Educators: share tasks, rubrics, and feedback to improve tutoring quality.
- Researchers: extend datasets, add evals, or submit fine‑tuned adapters.
- Partners: contact us for pilots, evaluations, or deployments.
Teach boldly. Build openly. 🐾
|