Be like a Goldfish, Donβt Memorize! Mitigating Memorization in Generative LLMs
Meta info.
- Authors: Abhimanyu Hans, Yuxin Wen, Neel Jain, John Kirchenbauer, Hamid Kazemi, Prajwal Singhania, Siddharth Singh, Gowthami Somepalli, Jonas Geiping, Abhinav Bhatele, Tom Goldstein
- Paper: https://arxiv.org/pdf/2406.10209
- Affiliation: University of Maryland
- Published: June 14, 2024
- Code: https://github.com/ahans30/goldfish-loss
TL; DR
causal language modeling objective λμ Goldfish Loss μ μ, μκΈ°λλ‘ μμ±ν΄λ΄λ λ°©μ μν




Suggestions
- νμ΅ν λ loss κ³μ°μμ μμ token subsetμ μ μΈνκ³ κ³μ° β long sequence μ¬μμ±νλ μ΄μ λ°©μ§
- νμ΅λ νΉμ ν ν°μ μ§λμΉκ² μ§μ€νλ λ¬Έμ μν
- subset μ νμ΄ μ±λ₯μ μ§μ μ μΈ μν₯
- G masking - λ Όλ¬Έμμλ λ¨μνκ² λ§€ k-th ν ν° λλ
- downstream task μ±λ₯ νλ½ κ±°μ μμ
- μλ§ standard λ°©μ λλΉ μΆ©λΆν νμ΅μ μκ°μ΄ λ μ€λ걸릴 건 μλͺ ν μ¬μ€β¦