LoRA adapter for Gemma 4 E2B (4-bit Unsloth base), fine-tuned for Socratic K–12 tutoring in realistic student register. Part of Canis Teach for the Gemma 4 Good hackathon (Future of Education · Unsloth · llama.cpp).

Trained on the full 51,870 multi-turn split of CanisAI/teach-r3-multilingual (dialogue field, default config). A separate adapter was trained on the 161k single-turn export; this repository is the 51k multi-turn publish.

Model details

Field	Value
Developed by	CanisAI (Marko Nedilko)
Adapter type	LoRA (PEFT / QLoRA)
Base model	`unsloth/gemma-4-E2B-unsloth-bnb-4bit`
Training framework	Unsloth + TRL
LoRA config	r=16, alpha=16, dropout=0
Training data	51,870 multi-turn dialogues
Data generation	Gemma 4 26B A4B-IT on DGX Spark + Canis.lab
Training hardware	RTX 4080 Super (full run impractical) → rented A6000, ~12 h
Languages	en, de, uk, fr, es, it (dataset coverage)
Code	https://github.com/crasyK/canis-gemma4good
Live Studio demo	https://canis.appwrite.network
Video	https://www.youtube.com/watch?v=QbxPs0jLiZY
License	Apache-2.0 (adapter). Base + derivatives: Gemma Terms of Use

Training run telemetry (loss CSVs) was not retained. The training config in the public repo is reproducible.

Intended use

Local / on-device Socratic tutoring (llama-server + Canis CLI)
Messy, short, informal student messages (slang, typos, fragments)
Research on small specialised tutors vs general chatbots

Out of scope: medical or legal advice, high-stakes assessment, unsupervised classroom deployment.

Training data

CanisAI/teach-r3-multilingual

Config	~Rows	Role
`default`	51,870	This adapter — multi-turn `dialogue`
`adapted-hybrid`	4,958	Adaption-enhanced hybrid slice
`adapted-hybrid-flat`	4,958	Flat single-turn variant
`chat-pilot-source`	300	Small adaptation experiment
`chat-pilot-adapted`	298	Small chat-column pilot — not this adapter’s train set

Gemma 4 on Gemma 4: 26B for synthetic data generation, E2B for this student-facing adapter.

How to use

Transformers + PEFT

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base = "unsloth/gemma-4-E2B-unsloth-bnb-4bit"
adapter = "CanisAI/teach-multilingual-gemma-4-e2b-r3"

tok = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter)

messages = [
    {"role": "system", "content": "You are a Socratic K-12 tutor. Do not give the final answer directly."},
    {"role": "user", "content": "hey kannst du mir einfach die lösung für aufgabe 3 schicken"},
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
out = model.generate(**tok(text, return_tensors="pt").to(model.device), max_new_tokens=256, temperature=0.7)
print(tok.decode(out[0], skip_special_tokens=True))

See model/load_adapter.py in the submission repo.

Unsloth (faster)

python

from unsloth import FastModel

model, tokenizer = FastModel.from_pretrained(
    model_name="CanisAI/teach-multilingual-gemma-4-e2b-r3",
    max_seq_length=2048,
)

Local deployment (llama.cpp / Ollama)

Merge or serve with llama.cpp / Ollama. See model/inference_ollama.md and cli/ in the submission repo.

Training procedure

Method: QLoRA (4-bit) via PEFT + Unsloth, supervised fine-tuning (SFT)
Base: unsloth/gemma-4-E2B-unsloth-bnb-4bit
LoRA: r=16, alpha=16
Data: 51,870 dialogues, default config, dialogue field
Hardware: A6000, ~12 hours (after 4080 Super projected ~1 week for full run)
Code / configs: https://github.com/crasyK/canis-gemma4good/tree/main/training

Evaluation (honest)

No formal benchmark suite published for this adapter at submission time.
No classroom field trial of this R3 adapter or Canis Studio.
R1 research A/B (Qwen3-era ELMs; tools in lesson/, Canis paper) informed dataset design (e.g. short student messages, generalist vs math-only). That is not validation of this Gemma 4 checkpoint in production.
Re-run training/train.py for your own loss curves.

Limitations and risks

Small E2B model → limited factual depth; can confabulate.
Synthetic R3 data → generator biases; uneven per-language coverage.
Optimises tutoring style, not factual correctness — use teacher oversight and/or RAG.
Not a child-safety certification; test before school deployment.

Citation

bibtex

@software{canis_teach_r3_gemma4_2026,
  author  = {Nedilko, Marko},
  title   = {teach-multilingual-gemma-4-e2b-r3: Socratic LoRA for Gemma 4 E2B},
  year    = {2026},
  url     = {https://huggingface.co/CanisAI/teach-multilingual-gemma-4-e2b-r3},
  note    = {Canis Gemma 4 Good submission, github.com/crasyK/canis-gemma4good}
}

Acknowledgments

Google DeepMind (Gemma 4), Unsloth, llama.cpp / Ollama communities, Appwrite (Studio hosting), Ollama × NVIDIA (GTC Golden Ticket / DGX Spark for R3 generation), participants in the R1 research study (Canis paper).