Scholarlink

<p>Modern Multilingual Large Language Models are built on gigantic networks, and training them in constrained environments is not always feasible. Modern Multilingual Large Language Models are also resource-intensive, and training these networks is time-consuming. PM4DA2E (Performant Multilingual Modulated and Multiplexed Memory Distilled Model with Adaptive Activation Ensembles) outshines by overcoming the challenges of Natural Language Processing Multilingual Large Language Models. PM4DA2E provides an alternate solution that is reliable and better suitable for constraint setups by introducing amendments to the novel transformer-based architecture with distillation, modulation, multiplexing, adapters, memory, and activation ensembles. PM4DA2E, due to its structural transformations, retains 99% of knowledge with 50% reduced network size. PM4DA2E achieves 51% inference speedup and 47% less training time on the XNLI (Cross-lingual Natural Language Inference) dataset / XNLI evaluation metrics (Accuracy and/or F1-Score) compared to competing Multilingual Large Language Network counterparts prevalent in this era.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Performant Multilingual Modulated and Multiplexed Memory Distilled Model with Adaptive Activation Ensembles

  • 期刊论文

  • 发表日期：

    2025年7月14日

查看全文
  • Subrit Dikshit,
  • Rahul Dixit,
  • Ritu Tiwari,
  • Priyank Jain

摘要

Modern Multilingual Large Language Models are built on gigantic networks, and training them in constrained environments is not always feasible. Modern Multilingual Large Language Models are also resource-intensive, and training these networks is time-consuming. PM4DA2E (Performant Multilingual Modulated and Multiplexed Memory Distilled Model with Adaptive Activation Ensembles) outshines by overcoming the challenges of Natural Language Processing Multilingual Large Language Models. PM4DA2E provides an alternate solution that is reliable and better suitable for constraint setups by introducing amendments to the novel transformer-based architecture with distillation, modulation, multiplexing, adapters, memory, and activation ensembles. PM4DA2E, due to its structural transformations, retains 99% of knowledge with 50% reduced network size. PM4DA2E achieves 51% inference speedup and 47% less training time on the XNLI (Cross-lingual Natural Language Inference) dataset / XNLI evaluation metrics (Accuracy and/or F1-Score) compared to competing Multilingual Large Language Network counterparts prevalent in this era.