Sebastian Raschka

@rikiyatakehi @a1zhang @lateinteraction Really cool results btw! Qwen3 models are a bit tricky though, so we will have to see if that also works for other models and tasks. I.e., if you take the Qwen3 base model (before any fine-tuning) and only do 50 steps of GRPO, you already get like ~15% -> ~50% eval accuracy on https://t.co/e8G6kKzpjo