{"tasks":[{"slug":"c6-mode-choice-intercity","scope":"international","geography":"Sydney–Canberra–Melbourne, Australia","phase":4,"family":"individual","scores":[{"slug":"gemma","value":0.7239361904761905,"skill":1.4216451553140108,"n_valid":210,"name":"Gemma-4-26B (API, temp 0)","display":"0.724","protocol":"Gemma-4-26B (API, temp 0)","url":"/tasks/c6-mode-choice-intercity"},{"slug":"llama","value":0.757192380952381,"skill":-3.106848638300508,"n_valid":210,"name":"Llama-3.3-70B (API, temp 0)","display":"0.757","protocol":"Llama-3.3-70B (API, temp 0)","url":"/tasks/c6-mode-choice-intercity"},{"slug":"gptoss","value":null,"skill":null,"n_valid":208,"name":"gpt-oss-120B (API, temp 0)","display":null,"protocol":"gpt-oss-120B (API, temp 0)","url":"/tasks/c6-mode-choice-intercity"}],"n_items":210,"n_test":210,"metric":"Brier (sum over modes)","higher":false,"unit":"","baseline":0.7343764172335601,"baseline_display":"0.734","source":"Travel mode choice, Greene and Hensher (1997)","url":"/tasks/c6-mode-choice-intercity","version":"historical","human":null,"best":{"slug":"gemma","value":0.7239361904761905,"skill":1.4216451553140108,"n_valid":210,"name":"Gemma-4-26B (API, temp 0)","display":"0.724","protocol":"Gemma-4-26B (API, temp 0)","url":"/tasks/c6-mode-choice-intercity"},"leaders":[{"slug":"gemma","value":0.7239361904761905,"skill":1.4216451553140108,"n_valid":210,"name":"Gemma-4-26B (API, temp 0)","display":"0.724","protocol":"Gemma-4-26B (API, temp 0)","url":"/tasks/c6-mode-choice-intercity"}],"status":"evaluated","title":"Predict individual mode choice on intercity trips"}],"aggregation":"Per task and version only; no cross-task average"}