High-capability language models can achieve substantial benchmark improvements from as few as 100 pristine, long-horizon data points