Understanding and Improving Noisy Embedding Techniques in Instruction Finetuning
The paper discusses advancements in instructional fine-tuning techniques, particularly focusing on noisy embeddings. It introduces a new method called SymNoise, which utilizes symmetric noise and significantly improves performance over existing methods. The findings suggest that further research into noise-based strategies in language model fine-tuning is essential.
- ▪The study analyzes the effectiveness of uniform and Gaussian noise in embedding techniques for instruction fine-tuning.
- ▪SymNoise, the new fine-tuning method, achieved a score of 69.04% on AlpacaEval, surpassing the previous state-of-the-art method, NEFTune.
- ▪The research indicates that symmetric noisy embeddings can enhance the model's performance by regulating local curvature more effectively.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.23171 (cs) [Submitted on 22 May 2026] Title:Understanding and Improving Noisy Embedding Techniques in Instruction Finetuning Authors:Abhay Yadav View a PDF of the paper titled Understanding and Improving Noisy Embedding Techniques in Instruction Finetuning, by Abhay Yadav View PDF HTML (experimental) Abstract:Recent advancements in instructional fine-tuning have injected noise into embeddings, with NEFTune (Jain et al., 2024) setting benchmarks using uniform noise. Despite NEFTune's empirical findings that uniform noise outperforms Gaussian noise, the reasons for this remain unclear. This paper aims to clarify this by offering a thorough analysis, both theoretical and empirical, indicating comparable performance among these noise types.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
