This work presents a novel architecture for building Retrieval-Augmented Generation (RAG) systems to improve Question Answering (QA) tasks from a target corpus. Large Language Models (LLMs) have revolutionized the analyzing and generation of human-like text. These models rely on pre-trained data and lack real-time updates unless integrated with live data tools. RAG enhances LLMs by integrating online resources and databases to generate contextually appropriate responses. However, traditional RAG still encounters challenges like information dilution and hallucinations when handling vast amounts of data. Our approach addresses these challenges by converting corpora into a domain-specific dataset and RAG architecture is constructed to generate responses from the target document. We introduce QuIM-RAG (Question-to-question Inverted Index Matching), a novel approach for the retrieval mechanism in our system. This strategy generates potential questions from document chunks and matches these with user queries to identify the most relevant text chunks for generating accurate answers. We have implemented our RAG system on top of the open-source Meta-LLaMA3-8B-instruct model by Meta Inc. that is available on Hugging Face. We constructed a custom corpus of 500+ pages from a high-traffic website accessed thousands of times daily for answering complex questions, along with manually prepared ground truth QA for evaluation. We compared our approach with traditional RAG models using BERT-Score and RAGAS, state-of-the-art metrics for evaluating LLM applications. Our evaluation demonstrates that our approach outperforms traditional RAG architectures on both metrics.

Large Language Models (LLMs) typically have billions of parameters and are thus often difficult to interpret in their operation. Such black-box models can pose a significant risk to safety when trusted to make important decisions. The lack of interpretability of LLMs is more related to their sheer size, rather than the complexity of their individual components. The TARS method for knowledge removal (Davies et al 2024) provides strong evidence for the hypothesis that that linear layer weights which act directly on the residual stream may have high correlation with different concepts encoded in the residual stream. Building upon this, we attempt to decode neuron weights directly into token probabilities through the final projection layer of the model (the LM-head). Firstly, we show that with Llama 3.1 8B we can utilise the LM-head to decode specialised feature neurons that respond strongly to certain concepts, with examples such as "dog" and "California". This is then confirmed by demonstrating that these neurons can be clamped to affect the probability of the concept in the output. This extends to the fine-tuned assistant Llama 3.1 8B instruct model, where we find that over 75% of neurons in the up-projection layers have the same top associated token compared to the pretrained model. Finally, we demonstrate that clamping the "dog" neuron leads the instruct model to always discuss dogs when asked about its favourite animal. Through our method, it is possible to map the entirety of Llama 3.1 8B's up-projection neurons in less than 15 minutes with no parallelization.

A critical part of creating code suggestion systems is the pre-training of Large Language Models on vast amounts of source code and natural language text, often of questionable origin or quality. This may contribute to the presence of bugs and vulnerabilities in code generated by LLMs. While efforts to identify bugs at or after code generation exist, it is preferable to pre-train or fine-tune LLMs on curated, high-quality, and compliant datasets. The need for vast amounts of training data necessitates that such curation be automated, minimizing human intervention. We propose an automated source code autocuration technique that leverages the complete version history of open-source software projects to improve the quality of training data. This approach leverages the version history of all OSS projects to identify training data samples that have been modified or have undergone changes in at least one OSS project, and pinpoint a subset of samples that include fixes for bugs or vulnerabilities. We evaluate this method using The Stack v2 dataset, and find that 17% of the code versions in the dataset have newer versions, with 17% of those representing bug fixes, including 2.36% addressing known CVEs. The deduplicated version of Stack v2 still includes blobs vulnerable to 6,947 known CVEs. Furthermore, 58% of the blobs in the dataset were never modified after creation, suggesting they likely represent software with minimal or no use. Misidentified blob origins present an additional challenge, as they lead to the inclusion of non-permissively licensed code, raising serious compliance concerns. By addressing these issues, the training of new models can avoid perpetuating buggy code patterns or license violations. We expect our results to inspire process improvements for automated data curation, with the potential to enhance the reliability of outputs generated by AI tools.

Quantized training of Large Language Models (LLMs) remains an open challenge, as maintaining accuracy while performing all matrix multiplications in low precision has proven difficult. This is particularly the case when fine-tuning pre-trained models, which often already have large weight and activation outlier values that render quantized optimization difficult. We present HALO, a novel quantization-aware training approach for Transformers that enables accurate and efficient low-precision training by combining 1) strategic placement of Hadamard rotations in both forward and backward passes, to mitigate outliers during the low-precision computation, 2) FSDP integration for low-precision communication, and 3) high-performance kernel support. Our approach ensures that all large matrix multiplications during the forward and backward passes are executed in lower precision. Applied to LLAMA-family models, HALO achieves near-full-precision-equivalent results during fine-tuning on various tasks, while delivering up to 1.31x end-to-end speedup for full fine-tuning on RTX 4090 GPUs. Our method supports both standard and parameter-efficient fine-tuning (PEFT) methods, both backed by efficient kernel implementations. Our results demonstrate the first practical approach to fully quantized LLM fine-tuning that maintains accuracy in FP8 precision, while delivering performance benefits.

Decoding human activity from EEG signals has long been a popular research topic. While recent studies have increasingly shifted focus from single-subject to cross-subject analysis, few have explored the model's ability to perform zero-shot predictions on EEG signals from previously unseen subjects. This research aims to investigate whether deep learning methods can capture subject-independent semantic information inherent in human EEG signals. Such insights are crucial for Brain-Computer Interfaces (BCI) because, on one hand, they demonstrate the model's robustness against subject-specific temporal biases, and on the other, they significantly enhance the generalizability of downstream tasks. We employ Large Language Models (LLMs) as denoising agents to extract subject-independent semantic features from noisy EEG signals. Experimental results, including ablation studies, highlight the pivotal role of LLMs in decoding subject-independent semantic information from noisy EEG data. We hope our findings will contribute to advancing BCI research and assist both academia and industry in applying EEG signals to a broader range of applications.

The rising demand for generative large language models (LLMs) poses challenges for thermal and power management in cloud datacenters. Traditional techniques often are inadequate for LLM inference due to the fine-grained, millisecond-scale execution phases, each with distinct performance, thermal, and power profiles. Additionally, LLM inference workloads are sensitive to various configuration parameters (e.g., model parallelism, size, and quantization) that involve trade-offs between performance, temperature, power, and output quality. Moreover, clouds often co-locate SaaS and IaaS workloads, each with different levels of visibility and flexibility. We propose TAPAS, a thermal- and power-aware framework designed for LLM inference clusters in the cloud. TAPAS enhances cooling and power oversubscription capabilities, reducing the total cost of ownership (TCO) while effectively handling emergencies (e.g., cooling and power failures). The system leverages historical temperature and power data, along with the adaptability of SaaS workloads, to: (1) efficiently place new GPU workload VMs within cooling and power constraints, (2) route LLM inference requests across SaaS VMs, and (3) reconfigure SaaS VMs to manage load spikes and emergency situations. Our evaluation on a large GPU cluster demonstrates significant reductions in thermal and power throttling events, boosting system efficiency.

Vision-Language Models (VLMs) have recently experienced significant advancements. However, challenges persist in the accurate recognition of fine details within high resolution images, which limits performance in multiple tasks. This work introduces Pheye, a novel architecture that efficiently processes high-resolution images while training fewer parameters than similarly sized VLMs. Notably, Pheye achieves a high efficiency while maintaining strong performance, particularly in tasks that demand fine-grained image understanding and/or the handling of scene-text.

The machine learning and data science community has made significant while dispersive progress in accelerating transformer-based large language models (LLMs), and one promising approach is to replace the original causal attention in a generative pre-trained transformer (GPT) with \emph{exponentially decaying causal linear attention}. In this paper, we present LeetDecoding, which is the first Python package that provides a large set of computation routines for this fundamental operator. The launch of LeetDecoding was motivated by the current lack of (1) clear understanding of the complexity regarding this operator, (2) a comprehensive collection of existing computation methods (usually spread in seemingly unrelated fields), and (3) CUDA implementations for fast inference on GPU. LeetDecoding's design is easy to integrate with existing linear-attention LLMs, and allows for researchers to benchmark and evaluate new computation methods for exponentially decaying causal linear attention. The usage of LeetDecoding does not require any knowledge of GPU programming and the underlying complexity analysis, intentionally making LeetDecoding accessible to LLM practitioners. The source code of LeetDecoding is provided at \href{https://github.com/Computational-Machine-Intelligence/LeetDecoding}{this GitHub repository}, and users can simply install LeetDecoding by the command \texttt{pip install leet-decoding}.

One of the ways to do mathematical analysis of the capabilities and limitations of the transformer architecture [8] is to study formal languages, recognizable by them. Namely, for a given formal language L, we study, if there exists a choice of parameters in the transformer architecture, for which words from L are accepted and words not from L are rejected by the resulting transformer. A seminal work of Hahn [5] performed such analysis for a number of formal languages, including the parity language, consisting of binary words with even number of 1s. Hahn have shown that transformers, recognizing this language, must have low confidence, partially explaining an empirically observed struggle of transformers in learning this language [1, 4]. However, this does not exclude the possibility of existence of a theoretical solution for this language.

With the expansion of neural networks, such as large language models, humanity is exponentially heading towards superintelligence. As various AI systems are increasingly integrated into the fabric of societies-through recommending values, devising creative solutions, and making decisions-it becomes critical to assess how these AI systems impact humans in the long run. This research aims to contribute towards establishing a benchmark for evaluating the sentiment of various Large Language Models in socially importan issues. The methodology adopted was a Likert scale survey. Seven LLMs, including GPT-4 and Bard, were analyzed and compared against sentiment data from three independent human sample populations. Temporal variations in sentiment were also evaluated over three consecutive days. The results highlighted a diversity in sentiment scores among LLMs, ranging from 3.32 to 4.12 out of 5. GPT-4 recorded the most positive sentiment score towards AGI, whereas Bard was leaning towards the neutral sentiment. The human samples, contrastingly, showed a lower average sentiment of 2.97. The temporal comparison revealed differences in sentiment evolution between LLMs in three days, ranging from 1.03% to 8.21%. The study's analysis outlines the prospect of potential conflicts of interest and bias possibilities in LLMs' sentiment formation. Results indicate that LLMs, akin to human cognitive processes, could potentially develop unique sentiments and subtly influence societies' perceptions towards various opinions formed within the LLMs.