README
Quiz Buzzer AI ⚡ — 早押しクイズAI
A head-to-head competitive buzz-quiz (早押しクイズ) where a human plays against an AI under equal conditions. Built for the HF Build Small Hackathon (total params ≤ 32B).
早押しクイズ (hayaoshi quiz) is a Japanese competitive-quiz format: the question is read aloud and players race to buzz in as early as they dare — buzzing too early on the lead-in clue is a costly false start. This project is specialized for Japanese; questions are in 日本語.
Links
- 💻 Source code (GitHub): https://github.com/YUGOROU/quiz-ai
- 🧠 Answering model: https://huggingface.co/YUGOROU/quiz-main-gemma-merged
- 🔔 Buzz-timing model: https://huggingface.co/YUGOROU/quiz-buzz-reg-1.2bjp-merged
How it works (two fine-tuned models, ≤ 32B total)
| Role | Model | Job |
|---|---|---|
| Buzz timing | YUGOROU/quiz-buzz-reg-1.2bjp-merged (LFM2.5-1.2B + regression head) | Reads the question char-by-char, emits a confidence; buzzes when conf ≥ θ. ~9 ms/char. |
| Answering | YUGOROU/quiz-main-gemma-merged (gemma-4-26B-A4B SFT) | From the partial question at buzz time, reasons (<think>…</think>) and answers. |
Both are fine-tuned on a quiz-grammar corpus derived from AI王 / JAQKET (corpus private; weights + inference code only). Total ≈ 27.2B params.
Architecture
- ZeroGPU: one
@spaces.GPU(duration=120)window precomputes a whole match (buzz position, reasoning, answer, 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). - Frontend: the spectator "Big Screen" UI (1920×1080) is a custom React frontend mounted on the
Gradio app's FastAPI. It fetches the live match from
POST /api/round. - Scoring: correct =
1.0 + 0.5·(1 − buzzFrac)(earlier buzz → bigger reward); wrong =−1.5.
Endpoints
/— Big Screen live UI (Start match → AI plays in real time)/api/round— generate a match with the real models (ZeroGPU)/gradio— minimal control panel (language, buzz θ)
Env
QUIZ_MAIN_REPO/QUIZ_BUZZ_REPO— model repos (defaults above)QUIZ_THETA— buzz threshold (default 0.6; higher = more cautious = later buzz = higher accuracy)QUIZ_LOAD_4BIT=1— load the main model in 4bit (lighter)QUIZ_MOCK=1— frontend dev without GPU (returns a dummy match)
Attribution / license
Training data derived from AI王 (Project AIO) / JAQKET. 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. The demo question pool is not committed to GitHub; regenerate it locally with
build_aio_pool.py (downloads AI王 data/aio, CC BY-SA 4.0) + genre labelling, then place the
resulting questions_pool_ja.json next to app.py.
irodori_tts/ is vendored from Aratako/Irodori-TTS-500M-v3
(reads the question aloud) and retains its upstream license. The announcer reference voice
(announcer_ref.wav) is likewise not committed — supply your own reference clip.
Contributors
29 commits
Similar repos
README
Quiz Buzzer AI ⚡ — 早押しクイズAI
A head-to-head competitive buzz-quiz (早押しクイズ) where a human plays against an AI under equal conditions. Built for the HF Build Small Hackathon (total params ≤ 32B).
早押しクイズ (hayaoshi quiz) is a Japanese competitive-quiz format: the question is read aloud and players race to buzz in as early as they dare — buzzing too early on the lead-in clue is a costly false start. This project is specialized for Japanese; questions are in 日本語.
Links
- 💻 Source code (GitHub): https://github.com/YUGOROU/quiz-ai
- 🧠 Answering model: https://huggingface.co/YUGOROU/quiz-main-gemma-merged
- 🔔 Buzz-timing model: https://huggingface.co/YUGOROU/quiz-buzz-reg-1.2bjp-merged
How it works (two fine-tuned models, ≤ 32B total)
| Role | Model | Job |
|---|---|---|
| Buzz timing | YUGOROU/quiz-buzz-reg-1.2bjp-merged (LFM2.5-1.2B + regression head) | Reads the question char-by-char, emits a confidence; buzzes when conf ≥ θ. ~9 ms/char. |
| Answering | YUGOROU/quiz-main-gemma-merged (gemma-4-26B-A4B SFT) | From the partial question at buzz time, reasons (<think>…</think>) and answers. |
Both are fine-tuned on a quiz-grammar corpus derived from AI王 / JAQKET (corpus private; weights + inference code only). Total ≈ 27.2B params.
Architecture
- ZeroGPU: one
@spaces.GPU(duration=120)window precomputes a whole match (buzz position, reasoning, answer, 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). - Frontend: the spectator "Big Screen" UI (1920×1080) is a custom React frontend mounted on the
Gradio app's FastAPI. It fetches the live match from
POST /api/round. - Scoring: correct =
1.0 + 0.5·(1 − buzzFrac)(earlier buzz → bigger reward); wrong =−1.5.
Endpoints
/— Big Screen live UI (Start match → AI plays in real time)/api/round— generate a match with the real models (ZeroGPU)/gradio— minimal control panel (language, buzz θ)
Env
QUIZ_MAIN_REPO/QUIZ_BUZZ_REPO— model repos (defaults above)QUIZ_THETA— buzz threshold (default 0.6; higher = more cautious = later buzz = higher accuracy)QUIZ_LOAD_4BIT=1— load the main model in 4bit (lighter)QUIZ_MOCK=1— frontend dev without GPU (returns a dummy match)
Attribution / license
Training data derived from AI王 (Project AIO) / JAQKET. 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. The demo question pool is not committed to GitHub; regenerate it locally with
build_aio_pool.py (downloads AI王 data/aio, CC BY-SA 4.0) + genre labelling, then place the
resulting questions_pool_ja.json next to app.py.
irodori_tts/ is vendored from Aratako/Irodori-TTS-500M-v3
(reads the question aloud) and retains its upstream license. The announcer reference voice
(announcer_ref.wav) is likewise not committed — supply your own reference clip.
Similar repos
Contributors
29 commits