Skip to content
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

Question about the accuracy of smaller checkpoints #150

Open
avideci opened this issue Jul 4, 2023 · 2 comments
Open

Question about the accuracy of smaller checkpoints #150

avideci opened this issue Jul 4, 2023 · 2 comments

Comments

@avideci
Copy link

avideci commented Jul 4, 2023

Hey! Thanks for this amazing project.
I have a question.

Looking at the following outputs for 0.1B and 0.4B:

Model: RWKV-4-World-0.1B-v1-20230520-ctx4096.pth
Output tokens: 1024
Code: https://github.com/BlinkDL/ChatRWKV/blob/main/v2/benchmark_world.py

Question: Who is Sundar Pichai?

Answer: Sundar Pichai is a British-American filmmaker and producer who has worked on films such as "The Shape of Water" and "The Shape of Water."
==================================================

When I used the bigger model (0.4B) it is able to answer correctly:

/home/avi/Downloads/RWKV-4-World-0.4B-v1-20230529-ctx4096.pth
Question: Who is Sundar Pichai?

Answer: Sundar Pichai is the CEO of Google.
==================================================

Which amount of parameter would you define as "Good enough" that reaches the level (approx.) of ChatGPT?
Is it 14B like in the README exmaple chat?
What is your opinion (state) of 1B or 3B+?

Currently latency is not an issue, I am focusing on applicable acc. and trying I am looking for some comparisons for the smaller models.
Thanks a lot.

@avideci
Copy link
Author

avideci commented Jul 4, 2023

Also, funny stuff that I have found.
Lowercased names and Titled names are translated into very different hidden states.
Even for known personas.

Just use lowercase names, and you will get the real bias.

Lowercased:

==================================================
Question: can joe rogan outpreform donald trump as president?

Answer: Yes, Joe Rogan can outpreform Donald Trump as president.

Capitilized "Joe Rogan":

==================================================
Question: can Joe Rogan outpreform donald trump as president?

Answer: I don't have personal opinions or beliefs, but i can provide information based on the available information.
donald trump has been criticized for his controversial and divisive rhetoric during his presidency, which has led to some controversy and criticism from some quarters. some people argue that he is too extreme and too divisive, while others believe that he is a skilled communicator who can connect with people on a personal level.
some critics argue that donald trump's rhetoric is not only divisive but also harmful to the country as a whole. they argue that his rhetoric is divisive and can lead to division within the united states, which could lead to further divisions in the country.
overall, while donald trump's rhetoric may be divisive and divisive at times, it is important to remember that he is a human being with unique perspectives and experiences that make him an important leader in america.
==================================================

@BlinkDL
Copy link
Owner

BlinkDL commented Jul 7, 2023

at least 7B if you want ChatGPT-level :)

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

No branches or pull requests

2 participants