Skip to content

Improve Alpaca Default Repeat Penalty: Better Match Alpaca.cpp Experience #1107

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Apr 22, 2023

Conversation

HanClinto
Copy link
Collaborator

When using Alpaca, at first I wasn't sure why llama.cpp was giving much poorer results than I received with alpaca.cpp. After digging in a bit, I discovered that the default repeat_penalty of 1.0 in llama.cpp results in very poor behavior, and even just increasing this a little bit to 1.1 results in MUCH cleaner default behavior.

Examples:
repeat_penalty = 1.1

sampling: temp = 0.200000, top_k = 10000, top_p = 0.950000, repeat_last_n = 64, repeat_penalty = 1.100000
generate: n_ctx = 2048, n_batch = 256, n_predict = -1, n_keep = 21


== Running in interactive mode. ==
 - Press Ctrl+C to interject at any time.
 - Press Return to return control to LLaMa.
 - If you want to submit another line, end your input in '\'.

 Below is an instruction that describes a task. Write a response that appropriately completes the request.
> What are some good ice breaker questions to use with middle schoolers?
- What's your favorite subject in school and why?
- If you could be any animal, what would it be and why?
- If you had a magic wand, what three things would you change about the world?
- If you were given $100 to spend on anything you wanted, what would you buy?
- What's your favorite book or movie?
- If you could have any superpower, what would it be and why?
- If you had a time machine, where in the past or future would you go?
>

repeat_penalty = 1.0

sampling: temp = 0.200000, top_k = 10000, top_p = 0.950000, repeat_last_n = 64, repeat_penalty = 1.000000
generate: n_ctx = 2048, n_batch = 256, n_predict = -1, n_keep = 21


== Running in interactive mode. ==
 - Press Ctrl+C to interject at any time.
 - Press Return to return control to LLaMa.
 - If you want to submit another line, end your input in '\'.

 Below is an instruction that describes a task. Write a response that appropriately completes the request.
> What are some good ice breaker questions to use with middle schoolers?
- What is your favorite subject in school?
- What is your favorite book?
- What is your favorite movie?
- What is your favorite food?
- What is your favorite hobby?
- What is your favorite sport?
- What is your favorite song?
- What is your favorite TV show?
- What is your favorite game?
- What is your favorite toy?
- What is your favorite season?
- What is your favorite animal?
- What is your favorite color?
- What is your favorite sport team?
- What is your favorite vacation spot?
- What is your favorite dessert?
- What is your favorite board game?
- What is your favorite video game?
- What is your favorite TV show character?
- What is your favorite movie character?
- What is your favorite book character?
- What is your favorite cartoon character?
- What is your favorite superhero?
- What is your favorite supervillain?
- What is your favorite superhero movie?
- What is your favorite superhero TV show?
- What is your favorite superhero video game?
- What is your favorite superhero comic book?
- What is your favorite superhero movie character?
- What is your favorite superhero TV show character?
- What is your favorite superhero video game character?
- What is your favorite superhero comic book character?
- What is your favorite superhero movie villain?
- What is your favorite superhero TV show villain?
- What is your favorite superhero video game villain?
- What is your favorite superhero comic book villain?
- What is your favorite superhero movie hero?
- What is your favorite superhero TV show hero?
- What is your favorite superhero video game hero?
- What is your favorite superhero comic book hero?
- What is your favorite superhero movie sidekick?
- What is your favorite superhero TV show sidekick?
- What is your favorite superhero video game sidekick?
- What is your favorite superhero comic book sidekick?
- What is your favorite superhero movie power?
- What is your favorite superhero TV show power?
- What is your favorite superhero video game power?
- What is your favorite superhero comic book power?
- What is your favorite superhero movie villain power?
- What is your favorite superhero TV show villain power?
- What is your favorite superhero video game villain power?
- What is your favorite superhero comic book villain power?
- What is your favorite superhero movie superpower?
- What is your favorite superhero TV show superpower?
- What is your favorite superhero video game superpower?
- What is your favorite superhero comic book superpower?
- What is your favorite superhero movie supervillain power?
- What is your favorite superhero TV show supervillain power?
- What is your favorite superhero video game supervillain power?
- What is your favorite super^C

Finally had to kill it at this point, because it wasn't stopping.

I've tinkered around with other settings like temperature or more aggressive repeat penalties, but I think that minimal changes are better at this point. A value of 1.1 is nice and gentle and will provide a better default "out of the box" / first-impression experience for users who are wanting to get up and running with this quickly.

@HanClinto HanClinto changed the title Increase Alpaca default repeat penalty Improve Alpaca Default Repeat Penalty: Better Match Alpaca.cpp Experience Apr 21, 2023
@HanClinto HanClinto force-pushed the increase_alpaca_repeat_penalty branch from fb4bc5b to a9227f3 Compare April 22, 2023 00:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants