After finetuning in a different dataset, output is wrong

[felri]

I have this dataset data.zip

logs from finetune

example of answers, when the question is related to the new dataset (like the last question), the output is the input?

Instruction: Tell me about alpacas. Response: Alpacas are animals that live in South America. They are related to llamas and camels. Alpacas are smaller than llamas and camels. Alpacas live in the Andes Mountains in Peru, Chile, and Bolivia. Alpacas are raised for their wool. Alpaca wool is soft and warm. Alpaca wool is used to make sweaters, hats, socks, and blankets. Alpacas are also raised for their meat. Alpaca meat is similar to beef.

Instruction:

Tell me about llamas.

Instruction: Tell me about the president of Mexico in 2019. Response: The president of Mexico in 2019 is Andrés Manuel López Obrador.

Instruction:

Tell me about the president of Mexico in 2019.

Instruction: In Stardew Valley, what are the characteristics of Traveling Cart? Response: In Stardew Valley, what are the characteristics of Traveling Cart?

Instruction:

In Stardew Valley, what are the characteristics of Traveling Cart?

[T-Atlas]

It seems your format is a bit confusing, which I can't understand very well.

[jonheng]

I faced similar issues but setting peft==0.2.0 in the requirements fixed it for me

[weiddeng]

Hi @felri in your instruction fine tune data, I see each instruction is prefixed by "In Stardew Valley, " What is the rationale? Have you tried without the prefix? Thanks!