Fix Qwen3 thinking mode + increase max_new_tokens: training/eval_adapter.py 664fa8e verified piyush-mk commited on Apr 25