Machine Learning

1 readers

1 users here now

Community Rules:

Be nice. No offensive behavior, insults or attacks: we encourage a diverse community in which members feel safe and have a voice.
Make your post clear and comprehensive: posts that lack insight or effort will be removed. (ex: questions which are easily googled)
Beginner or career related questions go elsewhere. This community is focused in discussion of research and new projects that advance the state-of-the-art.
Limit self-promotion. Comments and posts should be first and foremost about topics of interest to ML observers and practitioners. Limited self-promotion is tolerated, but the sub is not here as merely a source for free advertisement. Such posts will be removed at the discretion of the mods.

founded 11 months ago

MODERATORS

communick@academy.garden

[R] Fine-tuning transformer-based models (alien.top)

submitted 10 months ago by pryxon_36@alien.top to c/machinelearning@academy.garden

2 comments fedilink hide all child comments

Hello,

I am currently working on my thesis, which focuses on elucidating fake news through humor. My objective is to fine-tune a transformer-based model for this purpose. I have a question:

If I initially fine-tune the model to generate humor (using a prompt like "tell me a joke" and providing an expected response in the form of a joke), and then fine-tune it again (using a prompt like "explain why this news is fake" and providing the expected response), will the final model be capable of responding effectively to a prompt like "explain why this is fake in a funny manner" if I proceed with this approach?

Or should I fine-tune the model specifically for the prompt "explain why this is fake in a funny way" and provide it with the expected response in a similar manner?

Has anyone come across a "problem" like this and if so what do u think its the best approach ?

Thank you for the help!

top 2 comments

sorted by: hot top controversial new old

[–] gunshoes@alien.top 1 points 10 months ago (1 children)

It would more than likely lose the ability to "make funny." The problem is called "catastrophic forgetting" and it comes up when pretrained models are fine tuned on downstream tasks. There's some literature that shows the original pretraining induces some bias (English BERT models fine-tuned on multilingual sets retain English syntax patterns) but more often than not the model purges ability to perform it's original task.

[–] RegisteredJustToSay@alien.top 1 points 10 months ago

There's some good recent papers on how to tackle this. My favourite paper on the topic is probably here: Robust fine-tuning of zero-shot models - arXiv https://arxiv.org/pdf/2109.01903

But tl;dr: a weighted average of fine-tuned weights and original weights through a manually chosen weight tend to greatly mitigate this problem. I was surprised the paper didn't get more attention when it came out, but oh well.