5 ms·
Show HN: Make a free 3.8B model as reliable as one 7× bigger at parsing data
https://github.com/pcoz/llm-feedback-control https://github.com/pcoz/llm-feedback-control
- babhishek21 5mo agoLooking at the results, I have some thoughts: 1. I understand the need to have the ceiling model be at a big enough factor to make for a good headline. But is it really fair to compare between two different family of models (phi3:mini vs mixtral:8x7b)? 2. The corpora is really small. Are the results here statistically significant?