2307.07099.pdf
arxiv.org/pdf/2307.07099.pdf
1 Users
0 Comments
11 Highlights
0 Notes
Tags
Top Highlights
evaluate quality of generated data via fine-tuning small lan- guage models
Chain-of-Thoughts Attribute Manipulation (CoTAM),
these methods usually prompt LLMs to wildly generate new examples without proper con- trol, which hinders the informativeness of gener- ated data and might induce spurious correlation
n Step 1, we di- rectly query LLMs to decompose sentences into multiple attributes that are independent of task- specific attributes
we instruct LLMs to output a guideline to switch task-specific attributes and maintain the others
econstruct the sentence based on the guideline from Step 2
data flipping
o change the visual at- tributes in facial images
ener- ate texts from a continuous latent space discretely
CoT and propose a methodology to write a sentence with the same attributes as the input sen- tence.
As NK-shot has the same number of training data as CoTAM but with human annotation, we expect NK-shot human annotation to be the upper bound of our method
Ready to highlight and find good content?
Glasp is a social web highlighter that people can highlight and organize quotes and thoughts from the web, and access other like-minded people’s learning.