4 ms·

Yup, this isn't a theoretical concern, but a major practical one. GPT models are known for memorizing their training data: https://towardsdatascience.com/openai

by ipsum2 5y ago

Yup, this isn't a theoretical concern, but a major practical one. GPT models are known for memorizing their training data: https://towardsdatascience.com/openai-gpt-leaking-your-data-2dfb9e22c1b2 https://towardsdatascience.com/openai-gpt-leaking-your-data-... Edit: Github mentions the issue here: https://docs.github.com/en/github/copilot/research-recitation https://docs.github.com/en/github/copilot/research-recitatio... and here: https://copilot.github.com/#faq-does-github-copilot-recite-code-from-the-training-set https://copilot.github.com/#faq-does-github-copilot-recite-c... though they neatly ignore the issue of licensing :)

visarga 5y ago
> GPT models are known for memorizing their training data Hash each function, store the hashes as a blacklist. Then you can ask the model to regenerate the function until it is copyright safe.
- ipsum2 5y ago
  What if it copies only a few lines, but not an entire function? Or the function name is different, but the code inside is the same?
  - proteal 5y ago
    If we could answer those questions definitively, we could also put lawyers out of a job. There’s always going to be a legal gray area around situations like this.
    - ipsum2 5y ago
      Matching on the abstract syntax tree might be sufficient, but might be complex to implement.
  - vlasev 5y ago
    You can probably tokenize the names so they become irrelevant. You can ignore non-functional whitespace, so that code C remains. Maybe one can hash all the training data D such that hash(C) is in hash(D). Some sort of Bloom filter...
bogwog 5y ago
That second link says the following: > We found that about 0.1% of the time, the suggestion may contain some snippets that are verbatim from the training set That's kind of a useless stat when you consider that the code it generates makes use of your existing variable/class/function names when adapting the code it finds. I'm not a lawyer, but I'm pretty sure I can't just bypass GPL by renaming some variables.
kitsune_ 5y ago
It's not just about regurgitating training data during a beam search, it's also about being a derivative work, which it clearly is in my opinion.

by ipsum2 5y ago

visarga 5y ago

> GPT models are known for memorizing their training data Hash each function, store the hashes as a blacklist. Then you can ask the model to regenerate the function until it is copyright safe.

ipsum2 5y ago
What if it copies only a few lines, but not an entire function? Or the function name is different, but the code inside is the same?
- proteal 5y ago
  If we could answer those questions definitively, we could also put lawyers out of a job. There’s always going to be a legal gray area around situations like this.
  - ipsum2 5y ago
    Matching on the abstract syntax tree might be sufficient, but might be complex to implement.
- vlasev 5y ago
  You can probably tokenize the names so they become irrelevant. You can ignore non-functional whitespace, so that code C remains. Maybe one can hash all the training data D such that hash(C) is in hash(D). Some sort of Bloom filter...

bogwog 5y ago

That second link says the following: > We found that about 0.1% of the time, the suggestion may contain some snippets that are verbatim from the training set That's kind of a useless stat when you consider that the code it generates makes use of your existing variable/class/function names when adapting the code it finds. I'm not a lawyer, but I'm pretty sure I can't just bypass GPL by renaming some variables.

kitsune_ 5y ago

It's not just about regurgitating training data during a beam search, it's also about being a derivative work, which it clearly is in my opinion.