In recent years, large language models (LLMs) have shown exceptional capabilities across various natural language processing (NLP) tasks. However, such impressive performance often comes with the trade-off of an increased parameter size, posing significant challenges for widespread deployment. Knowledge distillation (KD) provides a solution by transferring knowledge from a large teacher model to a smaller student model. In this paper, we explore the task-specific distillation of LLMs at the logit level. Our investigation reveals that the logits of fine-tuned LLMs exhibit a more extreme long-tail distribution than those from vision models, with hidden “noise” in the long tail affecting distillation performance. Furthermore, existing logits distillation methods often struggle to effectively utilize the internal ranking information from the logits. To address these, we propose the Bi-directional Logits Difference (BiLD) loss. The BiLD loss filters out the long-tail noise by utilizing only top-k teacher and student logits, and leverages the internal logits ranking information by constructing logits differences. To evaluate BiLD loss, we conduct comprehensive experiments on 13 datasets using two types of LLMs. Our results show that the BiLD loss, with only the top-8 logits, outperforms supervised fine-tuning (SFT), vanilla KL loss, and five other distillation methods from both NLP and CV fields.

Large language models (LLMs) have demonstrated exceptional performance across diverse natural language processing tasks. However, as the model size and the input sequence’s length increase, the linearly increasing key-value (KV) cache significantly degrades inference throughput. Therefore, grouped-query attention (GQA), as an alternative to multi-head attention (MHA), has been widely introduced into LLMs. In this work, we propose a cost-effective method for converting MHA into GQA with any compression ratio of KV heads. The key point of our method lies in the application of Procrustes analysis to the attention heads, which enhances the similarity among attention heads while preserving computational invariance, thereby improving the model’s post-training performance. Subsequently, we employ L₀ regularization to prune redundant parameters. The model after pruning can be adapted to the standard GQA framework. Experimental results show that our strategy can compress up to 87.5% KV heads of LLaMA2-7B model and 75% KV heads of Sheared-LLaMA-1.3B with acceptable performance degradation. Our code is released at https://github.com/fpcsong/mha2gqa.

Capturing emotions within a conversation plays an essential role in modern dialogue systems. However, the weak correlation between emotions and semantics brings many challenges to emotion recognition in conversation (ERC). Even semantically similar utterances, the emotion may vary drastically depending on contexts or speakers. In this paper, we propose a Supervised Prototypical Contrastive Learning (SPCL) loss for the ERC task. Leveraging the Prototypical Network, the SPCL targets at solving the imbalanced classification problem through contrastive learning and does not require a large batch size. Meanwhile, we design a difficulty measure function based on the distance between classes and introduce curriculum learning to alleviate the impact of extreme samples. We achieve state-of-the-art results on three widely used benchmarks. Further, we conduct analytical experiments to demonstrate the effectiveness of our proposed SPCL and curriculum learning strategy.