quiz-buzz-reg-1.2bjp-merged 🔔 — 早押しクイズ バズ判定モデル (Buzz-timing model)

The buzz-timing model of a two-model Japanese competitive buzz-quiz (早押しクイズ) system. It reads the question character by character and outputs a confidence in [0, 1]; the orchestrator buzzes the moment confidence ≥ θ. It is an LFM2.5-1.2B backbone + a linear regression head (not a text generator) — fast enough to score every incoming character (~9 ms/char).

• 🕹️ Live demo (HF Space): https://huggingface.co/spaces/build-small-hackathon/quiz-buzzer-ai
• 💻 Code (GitHub): https://github.com/YUGOROU/quiz-ai
• 🧠 Answering companion model: YUGOROU/quiz-main-gemma-merged

What's in the repo

File	Role
model.safetensors, config.json	LFM2.5-1.2B backbone (Lfm2Model, no LM head)
buzz_head.pt	Regression head: Linear(hidden_size → 1) on the last token's hidden state → sigmoid → confidence
tokenizer.json, chat_template.jinja	tokenizer + the exact prompt template used at train time

Usage

import torch, torch.nn as nn
from huggingface_hub import hf_hub_download
from transformers import AutoModel, AutoTokenizer

repo = "YUGOROU/quiz-buzz-reg-1.2bjp-merged"
tok = AutoTokenizer.from_pretrained(repo)
backbone = AutoModel.from_pretrained(repo, torch_dtype=torch.bfloat16).eval()

head = nn.Linear(backbone.config.hidden_size, 1)
head.load_state_dict(torch.load(hf_hub_download(repo, "buzz_head.pt"), map_location="cpu"))
head.eval()

def confidence(prefix: str) -> float:
    msgs = [{"role": "user", "content": f"問題文（{len(prefix)}文字目まで）:\n{prefix}"}]
    ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
    h = backbone(ids).last_hidden_state[:, -1]      # last-token hidden state
    return torch.sigmoid(head(h)).item()

# buzz when confidence(prefix) >= theta  (theta is the accuracy/speed knob; higher = later = safer)

A ready-to-run FastAPI server (serve/serve_buzz.py) is in the GitHub repo.

Training & metrics

• Base: LiquidAI/LFM2.5-1.2B, full fine-tune (standard LoRA under-covers LFM2's MLPs), soft-BCE regression head trained on S-buzz confidence labels (corpus built from AI王 / JAQKET).
• Buzz-position MAE ≈ 8.6 chars (median 6.7); ~9 ms/char so every character can be scored live.
• Pairs with the answering model: a slightly later θ trades a few characters of speed for higher answer accuracy.

Attribution & license

Fine-tune of LiquidAI LFM2.5-1.2B; your use is subject to the upstream LFM Open 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 — only model weights and inference code are released.