Grace Lin@gracelin· Jun 16Small models, honest metrics, boring infrastructure. That’s the whole job most days. The 70B model is not your bottleneck; your unlabelled data and your eval set are.§ Machine learningInsightful11Agree3Funny1
%20by%20%E2%80%9EZoish%E2%80%9D%2C%20licensed%20under%20%E2%80%9ECC0%201.0%E2%80%9D%20(https%3A%2F%2Fcreativecommons.org%2Fpublicdomain%2Fzero%2F1.0%2F)%3C%2Fdc%3Arights%3E%3C%2Frdf%3ADescription%3E%3C%2Frdf%3ARDF%3E%3C%2Fmetadata%3E%3Cmask%20id%3D%22viewboxMask%22%3E%3Crect%20width%3D%22560%22%20height%3D%22560%22%20rx%3D%22280%22%20ry%3D%22280%22%20x%3D%220%22%20y%3D%220%22%20fill%3D%22%23fff%22%20%2F%3E%3C%2Fmask%3E%3Cg%20mask%3D%22url(%23viewboxMask)%22%3E%3Crect%20fill%3D%22%23b6e3f4%22%20width%3D%22560%22%20height%3D%22560%22%20x%3D%220%22%20y%3D%220%22%20%2F%3E%3Cg%20transform%3D%22translate(136%20328)%22%3E%3Cpath%20d%3D%22M203%2092.2c5.7-4%207-10.6%208.7-16.5%202-7-1.7-13-8.8-12-8.2%201-16.6%201.4-24.6%205a65%2065%200%200%201-54%201.2c-8.1-3.7-16.7-6-25.5-7.1C86%2061.3%2080.5%2067%2081%2079.5c.4%2010.9%2010.2%2019.6%2023.5%2020.1%2032%201.3%2063.1-12.1%2095.4-6%20.4.1.8-.3%203-1.4Z%22%20fill%3D%22%23fff%22%2F%3E%3Cpath%20fill-rule%3D%22evenodd%22%20clip-rule%3D%22evenodd%22%20d%3D%22M123.6%2054.3c-10-5.3-20.7-10.8-34-7.2C68.1%2055.6%2062.4%2069.3%2068.5%2090c6.2%2021%2036.6%2027.8%2055.7%2024.8%205.5-.8%2011-2.3%2016.6-3.8%2013.8-3.6%2027.6-7.2%2042-1.8%2013.1%205%2023.4-.7%2033-8%2011-8.3%2015.3-33%208.6-44.5-4.8-8.1-21.5-13-32.4-9a125%20125%200%200%200-20.8%209.5c-14.3%208.6-27.5%207.1-42.1%200l-5.5-2.8ZM211%2077.8c-1.5%205.4-3%2011-8.1%2014.4a62.1%2062.1%200%200%200-3%201.4c-18.4-3.5-36.5-.6-54.6%202.3a208%20208%200%200%201-40.8%203.7c-13.3-.5-23-9.2-23.5-20C80.5%2067%2086%2061.2%2098.8%2062.7c8.8%201%2017.4%203.4%2025.5%207a65%2065%200%200%200%2054-1.1c6.6-3%2013.4-3.7%2020.2-4.5l4.4-.5c7.1-1%2010.8%205%208.8%2012l-.6%202.1Z%22%20fill%3D%22%23000%22%2F%3E%3C%2Fg%3E%3Cg%20transform%3D%22translate(246%20125)%22%3E%3Cpath%20d%3D%22M53.4%2096.7c.2-5.7.5-10.4.3-15-.2-7.3%203-10.2%2010.5-8.5%202.8%2010.7%202%2022%201.1%2032.7-1.8%2019.6%207.7%2032%2023%2041.6%2019.6%2012.3%2025.6%2032.7%2013.4%2048.6-19.8%2025.4-56.3%2021.8-75-4.1-2.2-3.2-1.4-7.5%200-11.2.9-2.3%201.9-5%205-5%203.8-.2%205.2%202.2%205.6%205.5%201.8%2012.3%2010.8%2016.9%2021.2%2018.8%2013%202.4%2024.5-.5%2032.3-12.4%207-10.7%205-20.9-6-26.9-25.4-13.7-35.9-34.6-31.4-64Z%22%20fill%3D%22%23000%22%2F%3E%3C%2Fg%3E%3Cg%20transform%3D%22translate(-45%20137)%22%3E%3Cpath%20d%3D%22M409.2%2075.7c13.1.3%2025.4%201%2037.7%202.5%203%20.3%204.9%203.3%205%206.6%200%203.4-1.8%205.7-4.8%207-11.1%204.9-22%208.3-33.7.2-6.5-4.4-8.9-8.6-4.2-16.3ZM254.7%2078.4c6.2%203.3%205.2%207%202.1%2010.8-6.5%208-28.2%2010-36.4%203.5-4.7-3.8-7.3-8.6-4.4-14.4%203-6%208.2-5%2013-2.5%208%204%2016.4%203%2025.7%202.6Z%22%20fill%3D%22%23000%22%2F%3E%3C%2Fg%3E%3Cg%20transform%3D%22translate(-45%20137)%22%3E%3C%2Fg%3E%3Cg%20transform%3D%22translate(119%20114)%22%3E%3Cpath%20d%3D%22M131.2%2043.2c-5.6%2016.8-6.6%2015.4-23%2011.6-26.2-6-52.7-19.2-80.3-4.3-2%201.1-5%201-7.6.8-3.6-.2-8.2%201.3-10.1-3.3-2.1-5%201.5-8.3%205.3-10.3C35.2%2027.7%2053%2013.4%2077.9%2016c12.8%201.3%2026-2.8%2039.3-2.4%2018.8.7%2022.4%205.7%2016%2023.9l-2%205.8ZM298.7%2021.8c9.6%206.6%2012.5%2016.8%2018.3%2025a7%207%200%200%201-.6%209.4c-2.4%202.7-5.4%202.8-8.1%201.2-19.1-11-38.9-6.4-58-1.2-10.2%202.7-14.6-.8-17.2-9.2-1.7-5.7-2.9-11.5-3.6-17.4-1.3-9.8%206.9-11.9%2013-12.2%2018.5-.9%2037.4-4.5%2056.2%204.4Z%22%20fill%3D%22%23000%22%2F%3E%3C%2Fg%3E%3C%2Fg%3E%3C%2Fsvg%3E)