nemotron-cascade-rl

【数据与研究】Train language models through sequential, domain-wise RL stages (RLHF → Instruction-Following → Math → Code → SWE) without catastrophic forgetting. Exploit policy-dependent training data distribution where previous behaviors persist when reward-relevant. 14B model surpasses DeepSeek-R1-0528 (671B) on LiveCodeBench.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准