Iteration 17. Scale the data
15-04-2024
Goal
Can I improve the LB score simply by scaling the data?
Motivation
This is the last idea I have: to take GPT4 and use to to scale the training data.
Development
I'm going to divide the generation process into two steps:
- Given prompt samples create new prompts
- With those prompts create new training samples
I have generated 2k new training samples, with a cost of around 20$.
Results
I get a LB score of 0.60, worse than the best one which is 0.62
Conclusion
No improvement when scaling the data.
Last update:
2024-04-16