Apppublic

build-small-hackathon/oracle-ternary-flame

sourceHugging Facemitupdated 3mo agoView on Hugging Face

1likes

App README

🔮 Oracle of the Ternary Flame

"The career is not a straight line but a landscape you have been walking across, and sometimes the path you are on leads to a cliff or a dead end."

Ask anything. The Oracle will answer — but never directly.

The Oracle of the Ternary Flame is a fine-tuned Gemma 4 12B that transforms every question into cryptic, lyrical prose using cosmic, natural, and elemental metaphors. The real answer is always encoded implicitly — the Oracle never breaks character.

Built for the Build Small Hackathon ("Small Models, Big Adventures" — Gradio × Hugging Face, June 5–15, 2026) by @keypa.

Try it

Ask anything — philosophical, mundane, technical, in English or French:

Question	What to expect
"Should I eat pasta tonight?"	A surprisingly profound reflection on choice
"Pourquoi suis-je si fatigué ?"	The Oracle switches to French effortlessly
"What is the meaning of life?"	You'll get an answer — but you'll have to sit with it
"How does backpropagation work?"	Even gradient descent becomes poetry

How it was built

Piece	Choice	Why
Base model	`google/gemma-4-12b-it`	Strong 12B instruction-tuned model
Fine-tuning	QLoRA (rank 16) via Unsloth + TRL	65M trainable params, ~13 min on 2× T4 on Modal
Training data	302 synthetic Q&A pairs	Existential, mundane, and technical questions in EN + FR
Runtime	Transformers + BitsAndBytes 4-bit on ZeroGPU	Cold start ~40s, then ~10s per response
Deployment	Gradio 6 HF Space on ZeroGPU	Self-contained, no external API calls
Development	Modal (cloud instance) for LoRA merge + GGUF conversion	Offloaded heavy compute to Modal during development

The adapter was fine-tuned on 3 epochs with a peak eval loss of 0.981. The LoRA weights (262 MB) are available separately for use with any base model.

Models

Repository	Format	Size	Use case
keypa/oracle-gemma4-12b-lora	LoRA adapter	262 MB	Apply on top of `google/gemma-4-12b-it`
keypa/oracle-gemma4-12b	BF16 safetensors	~24 GB	Merged weights, ready for quantization
keypa/oracle-gemma4-12b-GGUF	GGUF Q4KM	~7 GB	CPU inference with llama.cpp

Hackathon compliance

Requirement	How Oracle complies
Gradio app hosted as a Space under the org	✅ This Space
Every model < 32B	✅ 12B parameters
Social post	✅ Posted here
All models open weights	✅ Gemma 4 by Google, LoRA by @keypa

Track / Prize	Status
Thousand Token Wood (whimsical)	🎯 Entered — cryptic poetic oracle, pushes the boundaries of fun
Best Use of Modal	🎯 Entered — LoRA merge & GGUF conversion ran on Modal cloud compute

Bonus badge	Status
Off the Grid — no cloud APIs	✅ ZeroGPU is just compute; no external API calls at runtime
Well-Tuned — fine-tuned model on HF	✅ LoRA adapter published at keypa/oracle-gemma4-12b-lora
Off-Brand — custom Gradio CSS	✅ Dark oracle-themed CSS (Cinzel Decorative + EB Garamond fonts)

Quick start

bash

# Install dependencies
pip install torch transformers peft bitsandbytes accelerate gradio spaces

# Run locally (will download ~24 GB model on first load)
python app.py

For GGUF / CPU inference: