AlpacaDataReflect

An experiment to see if chatgpt can improve the output of the stanford alpaca dataset

Advantages

• Very detailed
• Sometimes more accurate
• More creative story telling and poem writing

Limitations

• Too verbose for tasks asking for limited information
• Listing more items than required
• Sometimes over-corrects original output
• Sometimes mentions "here is updated version"

TODO:

• Think of ways to improve outputs further (having chatgpt select output that follows instruction closely)
• Train LORA / finetune