SPPO: Sequence-Level PPO for Long-Horizon Reasoning Tasks

Implement Sequence-Level PPO to stabilize LLM training for long reasoning tasks without the overhead of heavy critic models. This addresses a major bottleneck in open-source reasoning model fine-tuning.

Suggested repo: sppo-llm

"Reasoning models that don't lose their train of thought—now with stable RL."

Estimated effort: 80h