LLMs aren't a good indicator of success here because an LLM trained on 80% of the data is just as good as one trained on 100%, assuming the type/category of data is distributed evenly. Proxies help when you do need to get access to 100% of the data including data behind social media loginwalls.
alt Hacker News
LLMs aren't a good indicator of success here because an LLM trained on 80% of the data is just as good as one trained on 100%, assuming the type/category of data is distributed evenly. Proxies help when you do need to get access to 100% of the data including data behind social media loginwalls.