CanisAI/teach-math-qwen3-4b-2507-r1-merged

Canis.teach — Qwen3‑4B Instruct (Math) — Merged

Merged full model (LoRA adapters applied to the base), ready for direct use with Transformers.

• —Base: Qwen/Qwen3-4B-Instruct-2507
• —Release: CanisAI/teach-math-qwen3-4b-2507-r1-merged
• —Project: Canis.teach, Learning that fits.
• —Tags: canis-teach, qwen3, education, lora-merged, transformers

What is this?

This repository contains a merged checkpoint: the LoRA adapters fine‑tuned on Math tutoring dialogues have been merged into the base model (Qwen/Qwen3‑4B‑Instruct‑2507). This allows you to load and run the model directly with Transformers (no PEFT merge step at runtime).

For lightweight adapters or Ollama-friendly quantized builds, see the “Related” section.

Quick usage (Transformers)

from transformers import AutoTokenizer, AutoModelForCausalLM

repo = "CanisAI/teach-math-qwen3-4b-2507-r1-merged"

tok = AutoTokenizer.from_pretrained(repo, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
    repo,
    device_map="auto",
    torch_dtype="auto"
)

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

Recommended decoding (for instruct-style usage):

• —temperature ≈ 0.7
• —top_p ≈ 0.8
• —top_k ≈ 20 Adjust to your needs.

Intended use

• —Subject‑aware tutoring for Math with didactic, step‑by‑step responses.
• —Suitable for educational prototypes, demonstrations, and research.
• —Built to “teach, not just answer”: stepwise hints, clarity, and rubric‑aligned structure.

Safety and limitations

• —Human oversight is required. The model may hallucinate or oversimplify.
• —For fact‑heavy tasks, consider Retrieval‑Augmented Generation (RAG) with curriculum sources.
• —Follow data privacy and compliance rules in your environment (e.g., school policies).

Training summary

• —Base model: Qwen/Qwen3-4B-Instruct-2507
• —Method: Supervised fine‑tuning with LoRA (Unsloth + TRL/PEFT), then merged to full weights
• —Data: Subject‑specific tutoring dialogues generated/curated via Canis.lab
• —Goal: Improve clarity, hints, and step-by-step pedagogy for Math

Note: Exact hyperparameters and logs are provided in the LoRA training pipeline (if published) or available on request.

Related

• —LoRA adapters (lightweight):
• —CanisAI/teach-math-qwen3-4b-2507-r1
• —Quantized GGUF for Ollama/llama.cpp:
• —CanisAI/teach-math-qwen3-4b-2507-r1-gguf
• —Base model:
• —Qwen/Qwen3-4B-Instruct-2507

License

• —Inherits the base model’s license. Review the base model terms before use.
• —Dataset licensing and any third‑party assets should be respected accordingly.

Acknowledgments

• —Qwen3 by Qwen team
• —Unsloth, TRL, PEFT, and Transformers for training/serving
• —Educators and contributors supporting Canis.teach

Learning that fits.