This repository provides LoRA adapters fine-tuned on Math tutoring dialogues. Apply these adapters to the base model to enable subject-aware, didactic behavior without downloading a full merged checkpoint.

The model is designed to teach, not just answer - providing step-by-step explanations, hints, and pedagogically structured responses.

For ready-to-run merged models or Ollama-friendly GGUF quantizations, see the "Related Models" section.

Quick Start

Installation

pip install transformers peft torch

Usage (LoRA)

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "CanisAI/teach-math-qwen3-4b-2507-r1"

tokenizer = AutoTokenizer.from_pretrained(base, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
    base, 
    device_map="auto",
    torch_dtype="auto"
)
model = PeftModel.from_pretrained(model, adapter)

# Example prompt
prompt = "Explain how to solve 2x + 1 = 5 step by step."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.8,
    top_k=20,
    do_sample=True
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Base Model: Qwen/Qwen3-4B-Instruct-2507
Training Method: Supervised Fine-Tuning (SFT) with LoRA
Framework: Unsloth + TRL/PEFT
Data: Canis.lab-curated Math tutoring dialogues
Target Modules: Query, Key, Value, Output projections
Rank: 16
Alpha: 32

Intended Use

Primary: Subject-aware tutoring for Math education
Applications: Educational prototypes, tutoring systems, research
Approach: Stepwise explanations, pedagogical hints, rubric-aligned responses
Target Audience: Students, educators, researchers

Model Behavior

The model is optimized for:

Clear, step-by-step explanations
Appropriate difficulty progression
Encouraging learning through hints rather than direct answers
Subject-specific pedagogical approaches
Maintaining educational standards and accuracy

Recommended Settings

For optimal tutoring behavior:

Temperature: 0.6-0.8
Top-p: 0.8-0.9
Top-k: 20-40
Max tokens: 256-512 (depending on complexity)

Safety and Limitations

Important Considerations:

Human oversight required for educational use
May occasionally hallucinate or oversimplify complex topics
For fact-critical applications, consider RAG with verified curriculum sources
Follow your institution's data privacy and AI usage policies
Not a replacement for qualified human instruction

Related Models

Type	Repository	Description
LoRA Adapters	`CanisAI/teach-math-qwen3-4b-2507-r1`	This repository (lightweight)
Merged Model	(Coming Soon)	Ready-to-use full model
GGUF Quantized	(Coming Soon)	Ollama/llama.cpp compatible
Dataset	`CanisAI/teach-math-dataset`	Training data

License

This model inherits the license from the base model (Qwen/Qwen3-4B-Instruct-2507). Please review the base model's license terms before use.

Citation

bibtex

@misc{canis-teach-math,
  title={Canis.teach Math Tutor},
  author={CanisAI},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/CanisAI/teach-math-qwen3-4b-2507-r1}}
}

Acknowledgments

Qwen Team for the excellent base model
Unsloth for efficient training tools
Hugging Face ecosystem (Transformers, PEFT, TRL)
Educators and contributors supporting the Canis.teach project

Canis.teach - Learning that fits.