[DeepSeekMath] Pushing the Limits of Mathematical Reasoning in Open LMs

(Apr 2024)

 

섬세한 training data processing 와 RL algorithm 의 innovation (GRPO) 을 기반으로 놀라운 mathematical reasoning 성능 향상을 보여주었다. 

 

RL phase의 effectiveness를 보여주고, 어떤 요소에 기인하는지를 analyze해서 향후 성공적인 RL training 방향을 제시해준다. 

 

https://arxiv.org/pdf/2402.03300

https://github.com/deepseek-ai/DeepSeek-Math

GitHub - deepseek-ai/DeepSeek-Math: DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

---

---

---

※ Math Pre-Training / Supervised Fine-Tuning: data curation, training, evaluation 잘 설명되어있다. 읽었지만, 분량 상 생략.. 

---

---

---

---

---

---

---

---

---

---