Code for "APTBench: Benchmarking Agentic Potential of Base LLMs During Pre-Training" - View it on GitHub