This project performs full fine-tuning on the Qwen3-0.6B language model to enhance its mathematical problem-solving and reasoning capabilities. Training was conducted exclusively on the OpenMathReasoning-mini dataset, and the model was optimized using the bfloat16 (bf16) data type. The unsloth/OpenMathReasoning-mini dataset was used. Each example was formatted in Chain-of-Thought (CoT) style, pairing math problems with step-by-step intermediate reasoning. Qwen3 base model weights were loaded via the unsloth library in bf16 precision. All layers were updated (fullfinetuning=True`) to adapt the model for mathematical reasoning. Leveraged the Hugging Face TRL library with the Supervised Fine-Tuning (SFT) approach. The model was trained to generate both correct answers and corresponding reasoning chains. The model’s reasoning capacity for math problems was significantly improved through single-dataset, full fine-tuning in bf16 precision. Outputs include both intermediate reasoning steps and final solutions, providing transparent and interpretable results. This project is licensed under the Apache License 2.0. See the LICENSE file for details.
Модальности:
Генерация текста
Области применения:
Математика Логика и рассуждение Диалог / чат
Задача: Генерация текста
Автор: suayptalha
Теги: qwen3, unsloth, trl, sft, math, reasoning, conversational, en
Лайков: 3 | Загрузок: 166
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
