YUGOROU/quiz-main-gemma-merged

VISIONConcurrency Cost:2Model Size:26BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 6, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

YUGOROU/quiz-main-gemma-merged is a 26 billion parameter instruction-tuned Gemma 4 model developed by YUGOROU, specifically designed as the answering component of a Japanese competitive buzz-quiz system. This model, built for the HF Build Small Hackathon, processes partial quiz questions at buzz time, performs internal reasoning using a mechanism, and generates short answers. It is optimized for rapid, accurate responses in a quiz context, achieving 62-74% accuracy at the buzz position.

Loading preview...

YUGOROU/quiz-main-gemma-merged: Japanese Buzz-Quiz Answering Model

This model is the answering component of a two-model Japanese competitive buzz-quiz (早押しクイズ) system, developed by YUGOROU for the HF Build Small Hackathon. It is designed to receive a partial quiz question at the moment a player "buzzes in" and then generate a concise answer.

Key Capabilities & Features

  • Specialized Quiz Answering: Fine-tuned specifically for the unique challenge of answering quizzes based on incomplete information.
  • Reasoning Mechanism: Utilizes an internal <think>…</think> process to reason through the partial question before formulating an answer. Disabling this reasoning significantly impacts accuracy.
  • Gemma 4 Base: Built upon the unsloth/gemma-4-26B-A4B (MoE, 26B total / 4B active) base model, leveraging its gemma-4-thinking chat template.
  • Companion Model: Works in conjunction with YUGOROU/quiz-buzz-reg-1.2bjp-merged, which handles the buzz-timing aspect.
  • Performance: Achieves approximately 62–74% accuracy when answering at the buzz position, depending on the buzz threshold.

Usage & Technical Details

Users should ensure that the model's generation is stopped not only by the <eos> token (ID 1) but also by the <turn|> token (ID 106) to prevent hallucination of new turns. The model was fine-tuned using Unsloth bf16 LoRA on a quiz-grammar corpus derived from AI王 / JAQKET datasets. The training data includes partial questions at statistically determined buzz positions, with the assistant generating <think>{reasoning}</think>{answer}.

Licensing & Attribution

This model is a fine-tune of Google Gemma 4 and is distributed under the Apache License 2.0. The training data is derived from AI王 (Project AIO) / JAQKET, with quiz questions copyrighted by 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. It is intended for non-commercial research use only, and redistribution of the dataset is prohibited.

Finetunes
1 models