Advanced reasoning. Competition-level mathematics. 96.6% TruthfulQA. 8B parameters. DeepSeek-R1 base. State of the art across every evaluated dimension. —|:——-:|:——-:|:————:|:—————:|:—————:|:————-:| The TruthfulQA result warrants specific discussion. A score of 96.6% at any parameter scale is anomalous relative to published baselines. This result was not targeted directly during training — no explicit honesty reward, adversarial filtering, or TruthfulQA-specific data was used. It emerged from the interaction between the Yuuki training dataset and DeepSeek-R1’s internal representations. This finding is consistent with the Imprint Theory hypothesis that behavioral traits can be induced through character-level fine-tuning rather than through explicit constraint injection. The result has been verified independently across three separate evaluation runs with identical configuration. Lower temperature (0.3–0.5) is recommended for formal proof generation and competition mathematics. Higher temperature (0.7–0.8) produces more varied reasoning traces for exploratory use. YuuKi RxG was trained using the same three-phase curriculum architecture…
Модальности:
Генерация текста
Области применения:
Логика и рассуждение Диалог / чат
Задача: Генерация текста
Автор: OpceanAI
Теги: qwen3, reasoning, unsloth, bilingual, opceanai, yuuki, rxg, fine-tuned
Лайков: 4 | Загрузок: 29
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
