James Vo

James Vo

The advancement of Large Language Models (LLMs) has significantly transformed the field of natural language processing, although the focus on English-centric models has created a noticeable research gap for specific languages, including Vietnamese. To address this issue, this paper presents vi-mistral-x, an innovative Large Language Model designed expressly for the Vietnamese language. It utilizes a unique method of continual pre-training, based on the Mistral architecture, which incorporates grouped-query attention and sliding window attention techniques. This model, vi-Mistral-X, marks a significant step forward in improving the understanding and generation of the Vietnamese language. It introduces an additional phase of continual pre-training, specifically adapted for Vietnamese, enhancing the model's capability in understanding complex language nuances and generating accurate, context-aware Vietnamese text. Through comprehensive testing on various benchmarks, vi-mistral-x has shown to outperform existing Vietnamese LLMs in several key areas, including text classification, question answering, and text generation. Particularly, in the Vietnamese Multitask Language Understanding (VMLU) benchmark, vi-mistral-x sets a new standard, outperforming other available models significantly. This paper highlights the critical role of continual pre-training in advancing language-specific LLMs and opens new avenues for the development of multilingual models. We aim for vi-mistral-x to not just be an important asset for processing the Vietnamese language but also to encourage more advancements in creating large language models for languages that are less represented.

James Vo

As Large Language Models (LLMs) scale to longer context windows, the computational cost of attention mechanisms, which traditionally grows quadratically with input length, presents a critical challenge for real-time and memory-constrained deployments. Existing sparse attention techniques have sought to reduce this complexity, but they often incur significant overhead or compromise accuracy, making them less practical for large contexts on mid-range hardware. In this paper, we introduce SparseAccelerate, a dynamic sparse attention method that adapts its sparsity patterns based on input characteristics, effectively flattening the attention complexity curve. Our approach is effective for input lengths starting at 16K tokens and scales efficiently up to 128K tokens on dual NVIDIA A5000 GPUs (24GB each). Experimental results show that SparseAccelerate achieves up to a 1.04x reduction in Time-To-First-Token (TTFT) latency at 32K tokens, while also providing substantial memory savings. These improvements yield practical gains for memory-intensive applications and long-context tasks that were previously infeasible with standard attention. Beyond latency reductions, SparseAccelerate fundamentally shifts the scaling trend, demonstrating the smallest TTFT growth gradient relative to context length among competing methods. Ongoing evaluations on diverse benchmarks confirm its scalability, positioning SparseAccelerate as a critical advancement toward efficient, real-time, and large-context LLM inference on accessible hardware.

James Vo

As large language models (LLMs) continue to advance, the need for precise and efficient evaluation metrics becomes more pressing. Traditional approaches, while informative, often face limitations in computational demands and interpretability. In this paper, we introduce a novel hybrid evaluation method that integrates two established techniques: entropy derived from covariance matrices and the Matrix Nuclear Norm (MNN). Our method begins by normalizing hidden states from LLMs, then computes the covariance matrix and MNN from these representations. We further calculate the entropy of the covariance matrix to capture uncertainty and redundancy in the model's outputs. By combining these metrics into a composite score, we offer a comprehensive evaluation framework that balances accuracy with computational efficiency. Additionally, our approach allows for flexibility in adjusting the weightings between entropy and MNN, tailoring the evaluation for different objectives. Through a series of experiments on various LLMs, we demonstrate the robustness and efficacy of our method, offering deeper insights into model performance. This work contributes to the ongoing development of LLM evaluation and opens avenues for future innovations in model assessment techniques.

James Vo

In this paper, we propose Transformer Layer Injection (TLI), a novel method for efficiently upscaling large language models (LLMs) while minimizing computational costs and maintaining model performance. Model scale is a key factor in enhancing the quality of machine learning models, and TLI addresses the challenge of scaling by reducing initial loss, minimizing fine-tuning requirements, and preserving model complexity. Our approach improves upon the conventional Depth Up-Scaling (DUS) technique by injecting new layers into every set of K layers, enabling hidden representations to pass through transformer blocks with minimal disruption. We compare TLI with existing approaches, including Mixture of Experts (MoE) and DUS, and validate its efficiency through experiments on small LLMs (LLama3 1B, 3B, and 8B). Results show that TLI achieves better initialization, requires fewer training steps, and delivers superior accuracy on tasks such as KoBEST and KMCQA, with models performing effectively even without additional training. TLI is demonstrated to be both data-efficient and cost-effective, significantly outperforming existing methods. Its scalability and simplicity make it a promising solution for upscaling transformer-based models, with potential applications in scaling models from 10B to 405B parameters.