Modelpublic

Gaurav-x111/sensei-origamiAI-v1

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes4downloads

Model Card

🦢 Sensei — OrigamiAI Teaching Model (v1)

Sensei is a fine-tuned origami teaching assistant built for the OrigamiAI project. It teaches users how to fold origami step by step, with three distinct personality modes.

v1 — baseline release. v2 in progress with improved accuracy.

Model Details

Property	Value
Base model	`Qwen/Qwen3-1.7B`
Fine-tuning method	LoRA (PEFT)
Training pairs	2,757
Origami models covered	118
Training epochs	5
Max sequence length	256 tokens
Version	v1

Evaluation (v1)

Evaluated using a custom origami-domain benchmark on 15 held-out step samples.

Metric	Score	Status
ROUGE-L	0.183	Expected for generative model
Knowledge score	0.500	Improving in v2
Repetition rate	0.000	✓ No looping
Personality distinctiveness	0.840	✓ Strong
Avg response time (CPU i5)	12.4s	—
Overall score	0.631	Good

Note: ROUGE-L is intentionally low — the model paraphrases rather than copies training data, which is correct behavior for a teaching assistant.

Personalities

Sensei supports three teaching styles injected via system prompt at runtime:

Avatar	Style	Prompt prefix
ZANG	Strict, precise master	`You are ZANG, a strict traditional origami master. Demand precision. Be brief and stern.`
ZENU	Playful, enthusiastic	`You are ZENU, an energetic fun origami friend! Use emojis, be enthusiastic!`
MOMO	Gentle, patient	`You are MOMO, a gentle patient origami teacher. Speak softly, never rush.`

Usage

Ollama (recommended for local use)

FROM Gaurav-x111/sensei-origami-v1

SYSTEM You are Sensei Kami, a wise origami master. Teach origami step by step. Keep answers to 1-2 sentences.

PARAMETER repeat_penalty 1.3
PARAMETER temperature 0.3
PARAMETER num_predict 80
PARAMETER num_thread 8
PARAMETER num_ctx 2048
PARAMETER top_k 20
PARAMETER stop "\n\n"
TEMPLATE """<|im_start|>system
{{ .System }}<|im_end|>
<|im_start|>user
/no_think {{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""

Python (transformers)

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Gaurav-x111/sensei-origami-v1")
tokenizer = AutoTokenizer.from_pretrained("Gaurav-x111/sensei-origami-v1")

prompt = "### Instruction:\nCrane step 3 done. Give encouragement.\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=80, temperature=0.3, do_sample=True)
print(tokenizer.decode(out[0], skip_special_tokens=True).split("### Response:")[-1].strip())

Training Data

—Source: 119 origami instruction pages scraped and labeled using meta/llama-3.2-11b-vision-instruct
—Format: Step-by-step fold instructions with fold type, direction, and location
—Augmented with: 60+ handcrafted encouragement Q&A pairs per avatar personality
—Pipeline: ml/scrape.py → ml/label.py → ml/train.py

Project

Part of OrigamiAI — a locally-runnable origami tutoring system with:

—Real-time hand + paper tracking (MediaPipe + OpenCV)
—3D origami simulator sync (Origami Simulator + WebSocket)
—Three Sensei avatar personalities
—Full offline operation — no cloud APIs

GitHub: Gaurav-x111/orikamiai

Roadmap

—v2 — 10 epochs, 512 token context, improved knowledge score target 0.75+
—v3 — full dataset expansion, multi-turn conversation support

License

Apache 2.0 — free to use, modify, and distribute.