Sabitlenmiş Tweet
📣📣 New research on AI safety
Language Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic
We are introducing a new approach to prevent LLM's compromised safety
Paper: arxiv.org/pdf/2402.11746…
Code and safety evaluation dataset: github.com/declare-lab/re…
Data: huggingface.co/datasets/decla…
Did you know enterprise adaptation of LLMs comes at the cost of compromised safety? If not, our research might be interesting to you.
We propose RESTA: Restoring Safety through Task Arithmetic.
👉 RESTA is a simple, fast, and effective approach that provides a no-cost solution to model safety realignment. At the core of RESTA, we add a homegrown safety vector to the model to bring back its safety while maintaining its task performance.
👉 Our evaluations of fine-tuned Llama-2 on CatQ (a multilingual safety evaluation benchmark we created) show a sharp drop in unsafety score from 33.57% to 12.17% in PEFT and from 22.16% to 4.34% in Full-FT averaged across fine-tuning domains.
👉 To gauge the effect of the safety vector beyond categories in CatQ, we evaluate RESTA on three existing safety evaluation benchmarks— HarmfulQ, AdevrsarialQ, and DangerousQ. We observe a reduction in the unsafety score from 18.59% to 5.14% in PEFT and from 9.16% to 1.55% in Full-FT when averaged across benchmark datasets and fine-tuning domains.
👉 The effectiveness of RESTA is evident across languages, as seen in the 26.2% reduction in PEFT and 21.37% reduction in Full-FT on the Vietnamese CatQ. Similar improvements are observed for CatQ in Chinese, with a reduction of 17.35% in PEFT and 24.54% in Full-FT.
For more interesting information, please have a look at our work at lnkd.in/gr9dNz5H.
Thanks to my great collaborators, @soujanyaporia and Duc Anh Do!
#NLProc #Safety #LLM @llm_sec @topofmlsafety @cohere @dair_ai and hoping @seb_ruder @seraphinagt @jeremyphoward @omarsar0 like it.
English
