MiniSymp2 / README.md
kalomaze's picture
Update README.md
f05fa86
metadata
license: apache-2.0

MiniSymp2 is A retrain of my MiniSymposium model attempt except with some more data and better practices.

  • added EOS tokens where they belong
  • made the prompt formats more diverse in the data so you could experiment / play with prompt format in context
  • added some new examples
  • measured loss curve to make sure I wasn't overfitting
  • used 8-bit lora instead of 4-bit qlora