Deepseek R1 Distill Llama 70B

We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrated remarkable performance on reasoning. With RL, DeepSeek-R1-Zero naturally emerged with numerous powerful and interesting reasoning behaviors. However, DeepSeek-R1-Zero encounters challenges such as endless repetition, poor readability, and language mixing. To a...

本研究推出了初代推理模型DeepSeek-R1-Zero与DeepSeek-R1。其中DeepSeek-R1-Zero是一款无需以监督微调（Supervised Fine-Tuning，SFT）作为前置训练步骤，通过大规模强化学习（Reinforcement Learning，RL）训练得到的模型，在各类推理任务中展现出了卓越的性能。借助强化学习，DeepSeek-R1-Zero自发涌现出了诸多强大且颇具趣味的推理行为。然而，DeepSeek-R1-Zero也面临着诸如无限重复、可读性欠佳以及语言混杂等挑战。To a...

搜集汇总

背景与挑战

背景概述

该数据集包含第一代推理模型DeepSeek-R1-Zero，该模型通过大规模强化学习训练展现出卓越的推理能力，但存在输出重复、可读性不佳及语言混杂等缺陷。