# engine (pinned by the competition kit) vllm==0.31.0 # data generation / training (installed next to the engine) accelerate datasets safetensors numpy huggingface_hub