Tag
This paper investigates instruction finetuning of DeepSeek-R1-8B using LoRA and NEFTune for financial named-entity recognition, achieving a micro-F1 of 0.912 and outperforming several baseline models.
This paper analyzes noisy embedding techniques for instruction fine-tuning, explains why uniform noise outperforms Gaussian, and introduces SymNoise, a symmetric noise method that significantly improves LLaMA-2-7B performance on AlpacaEval over NEFTune.