This paper surveys recent replication studies of DeepSeek-R1, focusing on Supervised Fine-Tuning (SFT) and Reinforcement Learning from Verifiable Rewards (RLVR). It details data construction, method design, and training procedures, offering insights and anticipating future research directions for reasoning language models.
The article introduces a new approach to language modeling called test-time scaling, which enhances performance by utilizing additional compute resources during testing. The authors present a method involving a curated dataset and a technique called budget forcing to control compute usage, allowing models to double-check answers and improve reasoning. The approach is demonstrated with the Qwen2.5-32B-Instruct language model, showing significant improvements on competition math questions.
Trained on a vast dataset comprising primarily GPT-4 generated data and supplemented with high-quality information from open datasets in the AI field, this model exhibits exceptional performance across various tasks. It introduces a novel SFT + DPO version, and for those who prefer a different approach, an SFT-only version is also made available
ChatQA, a new family of conversational question-answering (QA) models developed by NVIDIA AI. These models employ a unique two-stage instruction tuning method that significantly improves zero-shot conversational QA results from large language models (LLMs). The ChatQA-70B variant has demonstrated superior performance compared to GPT-4 across multiple conversational QA datasets.
