Fine-tuned from DeepScaleR-1.5B. Surpasses it on every benchmark. 1.5B parameters. RTX 4080. Three-phase curriculum training. —|:———:|:———:|:———:|:————:|:————-:|:—:| Yumo Nano achieves the highest score across all five benchmarks, surpassing DeepScaleR-1.5B — the model it was derived from — on every individual metric. The most significant improvement is on OlympiadBench (+2.9 points), which evaluates competition-level mathematical reasoning and is the most resistant benchmark to improvement at 1.5B scale. The improvement on Minerva Math (+2.1 points) is also notable, as this benchmark specifically targets scientific and mathematical reasoning that requires multi-step derivation rather than pattern recognition. For high-precision computation tasks, reduce temperature to 0.3–0.5. Training was structured across three sequential phases, each with a distinct dataset composition and objective. All phases draw from the same four sources in different proportions. Establish mathematical identity and conversational baseline. — Hendrycks Math: Levels 1–3 only. Competition-level capability (Levels 4–5) is inherited from DeepScaleR base weights and was not…
Модальности:
Генерация текста
Области применения:
Логика и рассуждение Диалог / чат
Задача: Генерация текста
Автор: YU-MO
Теги: gguf, qwen2, reasoning, unsloth, bilingual, opceanai, yumo, fine-tuned
Лайков: 4 | Загрузок: 38
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
