CanisAI

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. 🐾