English · 日本語

quiz-ai ⚡ — a Japanese competitive buzz-quiz LLM system

A head-to-head competitive buzz-quiz (早押しクイズ) where a human plays against an AI under equal conditions — no handicap. The question is revealed one character at a time and both sides race to buzz in as early as they dare; buzzing too early on the lead-in is a costly false start. Built for the HF Build Small Hackathon (total params ≤ 32B).

早押しクイズ (hayaoshi quiz) is a Japanese competitive-quiz format. This project is specialized for Japanese — questions are in 日本語, and the two models are fine-tuned on Japanese quiz grammar.

Demo

▶️ Try it live: https://huggingface.co/spaces/build-small-hackathon/quiz-buzzer-ai

How it works — two fine-tuned models (≤ 32B total)

Both are fine-tuned on a quiz-grammar corpus derived from AI王 / JAQKET (≈ 27.2B params total).

Architecture

• One GPU window = one match. On ZeroGPU a single @spaces.GPU(duration=120) call precomputes a whole match — buzz position, reasoning, answer and correctness for N questions. The frontend then plays it back smoothly, streaming the question as mock-STT at 22 chars/s while the human can buzz in live (Space / tap).
• Custom React frontend (the spectator "Big Screen", 1920×1080) is served by a Gradio app and fetches the live match from POST /api/round.
• Scoring: correct = 1.0 + 0.5·(1 − buzzFrac) (earlier buzz → bigger reward); wrong = −1.5.
• Speculative & async: the answering model starts reasoning before the buzz is committed; <think> reasoning is streamed so the AI answers in a few seconds — like a human pausing to think.
• TTS: Irodori-TTS reads the question aloud, synced to the character reveal.

Repository layout

quiz-ai/
├── docs/         Design docs (quiz-ai.md / corpus.md) and assets
├── src/          Corpus preprocessing, Phase-0 orchestrator, shared utilities
│   ├── qutils.py            normalization · scoring (is_correct) · LLM client · qid split
│   ├── annotate.py          Step 1: S-buzz annotation
│   ├── build_corpus{1,2}.py Step 2/3: buzz-model / answering-model corpora
│   ├── p0_orchestrator.py   Phase 0: asyncio speculative-inference orchestrator
│   └── buzz_client.py       conf ≥ θ buzz decision client
├── train/        Training & evaluation (run on Modal)
│   ├── sft.py / modal_sft.py       SFT for the answering & buzz models
│   ├── buzz_reg.py / buzz_rl.py    buzz regression head / single-model RL
│   ├── eval_knowledge.py           knowledge-ceiling & full/prefix accuracy
│   ├── eval_buzz.py                buzz-position MAE
│   └── e2e_modal.py                end-to-end (buzz → speculative answer → scoring)
├── serve/        Inference serving (buzz FastAPI / main vLLM)
├── bench/        Latency benchmarks (LFM2.5 decode; no question text)
└── space/        HF Space (ZeroGPU + Gradio) live demo

Models

Running

• Training / evaluation runs on Modal: uv run --with modal modal run train/….
• Live demo is the HF Space (ZeroGPU) under space/ — see space/README.md.
• All scripts assume uv run (do not call python3 directly).

Data & license

• Training data is derived from AI王 (Project AIO) / JAQKET. The quiz questions are not redistributed: the corpus and demo question pool (corpus/, annotated_questions.jsonl, questions_*.json, …) are git-ignored. Regenerate the demo pool locally with space/build_aio_pool.py (downloads AI王 data/aio, CC BY-SA 4.0) + src/label_genres.py.
• Only model weights + training/inference code are published, with attribution:

  Quiz questions © abc/EQIDEN実行委員会 / 株式会社キュービック / クイズ法人カプリティオ. Non-commercial research use only. No dataset redistribution.

• space/irodori_tts/ is vendored from Aratako/Irodori-TTS-500M-v3 and retains its upstream license.
• The answering model is a fine-tune of Google Gemma 4, which is released under the Apache License 2.0; the buzz model inherits the LFM Open License.

Acknowledgements

AI王 / Project AIO · JAQKET · Google Gemma · LiquidAI LFM2.5 · Irodori-TTS · Unsloth · the HF Build Small Hackathon.