Search Results for « Instruction Tuning »

The growing demand for natural language processing models capable of understanding and executing complex instructions has driven significant advancements in model fine-tuning techniques. The novel concept of instruction tuning, which involves fine-tuning pre-trained language models on meticulously curated instruction datasets, has shown remarkable promise in enhancing model performance. The research presented here focuses on applying instruction tuning to GPT2 (124M parameters) to improve its reasoning capabilities on the Multi-task Language Understanding (MMLU) dataset. By systematically curating a diverse set of tasks and corresponding instructions, and rigorously fine-tuning the model, significant improvements were achieved in key performance metrics, including accuracy, precision, recall, and F1-score. Experimental results demonstrated that the instruction-tuned GPT-2 model significantly outperformed the baseline GPT-2 and other stateof-the-art models, showcasing the effectiveness of the instruction tuning approach. The enhanced capacity of the model to follow detailed instructions led to more accurate and contextually relevant responses, showing the potential of this methodology to refine and augment the capabilities of pre-trained models. The comprehensive preparation of the instruction dataset and the iterative tuning process were critical factors in achieving these substantial performance gains. The study’s findings suggest that instruction tuning can be a powerful tool for optimizing the performance of language models across a variety of tasks and domains, provided that the instruction datasets are carefully curated and validated. The instruction tuning of GPT-2 (124M parameters) resulted in significant improvements in the model’s reasoning capabilities, as evidenced by the enhanced performance metrics on the MMLU dataset. The research highlights the potential of instruction tuning as an effective approach for refining pre-trained models and enhancing their applicability in complex and diverse scenarios. By demonstrating the substantial benefits of fine-tuning models on carefully prepared instruction datasets, the study provides valuable insights into the potential of this technique for further advancements in natural language processing.

Abstract To self-renew, stem cells must precisely balance proliferation and differentiation. Typically, this is achieved under feedback from the niche; yet many stem cells also possess an intrinsic self-renewal program that allows them to do so autonomously, as required. However, because self-renewal implies a stable equilibrium – in which the expected stem cell number neither increases nor decreases over time – this seems to require fine-tuning to a critical point. Here, we show that this is not the case: self-renewal can, in principle, be easily achieved without the need for extrinsic instruction, feedback or finetuning, by a simple ‘dimerization cycle’ that uses partitioning errors at cell division to reliably establish asymmetric divisions and perfectly balance symmetric divisions.

Advances in large language models (LLMs) have transformed the field of natural language processing and have enormous potential for social scientific analysis. We explore the application of LLMs to supervised text classification. As a case study, we consider stance detection and examine variation in predictive accuracy across different architectures, training regimes, and task specifications. We compare ten models ranging in size from 86 million to 1.7 trillion parameters and four distinct training regimes: prompt-based zero-shot learning; few-shot learning; fine-tuning; and instruction-tuning. The largest models generally offer the best predictive performance, but fine-tuning smaller models is a competitive solution due to their relatively high accuracy and low cost. For complex prediction tasks, instruction-tuned open-weights models can perform well, rivaling state-of-the-art commercial models. We provide recommendations for the use of LLMs for text classification in sociological research and discuss the limitations and challenges related to the use of these technologies.

Advances in large language models (LLMs) have transformed the field of natural language processing and have enormous potential for social scientific analysis. We explore the application of LLMs to supervised text classification. As a case study, we consider stance detection and examine variation in predictive accuracy across different architectures, training regimes, and task specifications. We compare ten models ranging in size from tens of millions to hundreds of billions of parameters and four distinct training regimes: prompt-based zero-shot learning and few-shot learning, fine-tuning using more training data, and instruction-tuning that combines prompting and training data. The largest models generally offer the best predictive performance even with little or no training examples, but fine-tuning smaller models is a competitive solution due to their relatively high accuracy and low cost. For complex prediction tasks, instruction-tuned open-weights models can perform well, rivaling state-of-the-art commercial models. We provide recommendations for the use of LLMs for text classification in sociological research and discuss the limitations and challenges related to the use of these technologies.

Advances in large language models (LLMs) have transformed the field of natural language processing and have enormous potential for social scientific analysis. We explore the application of LLMs to supervised text classification. As a case study, we consider stance detection and examine variation in predictive accuracy across different architectures, training regimes, and task specifications. We compare ten models ranging in size from tens of millions to hundreds of billions of parameters and four distinct training regimes: prompt-based zero-shot learning and few-shot learning, fine-tuning using more training data, and instruction-tuning that combines prompting and training data. The largest models generally offer the best predictive performance even with little or no training examples, but fine-tuning smaller models is a competitive solution due to their relatively high accuracy and low cost. For complex prediction tasks, instruction-tuned open-weights models can perform well, rivaling state-of-the-art commercial models. We provide recommendations for the use of LLMs for text classification in sociological research and discuss the limitations and challenges related to the use of these technologies.

Multimodal large language models (MLLMs), which can reason across different modalities, are opening up new frontiers in AI development that will soon enable complex reasoning with both visual and textual modalities. One application of multimodal MLLMs is Visual Question Answering (VQA) in healthcare; for example, a clinician can ask questions about medical images in natural language to receive detailed explanations. As suitably adapted for the modern age: "The art of medicine consists in amusing the patient while nature cures the disease - but in the age of AI, assisting the doctor while learning from the data." - Inspired by Voltaire, adapted for modern AI healthcare This research examines the efficacy of fine-tuning multimodal (vision and language) foundational models for performing medical visual question answering. This study aims to evaluate how effectively these models can interpret and respond to medical queries based on visual inputs, ultimately enhancing diagnostic accuracy and patient care. By leveraging the strengths of both vision and language processing, the research seeks to advance the capabilities of AI in medical settings. The author studies three fine-tuned variants derived from the baseline LLaVA-Med architecture: a caption-only baseline trained to generate global descriptions of medical images; a model that uses instruction tuning to better answer diagnostic and region-specific questions, and a variant that incorporates Alpha-CLIP with LLaVA-Med to allow spatially targeted understanding by conditioning responses on user-defined regions of interest (ROIs). All variants are adapted utilizing Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning (PEFT) technique that facilitates minimal computational overhead. The results show that the foundational LLaVA-Med model, which was trained only on image-caption pairs, performs fairly well with BLEU (0.09) and ROUGE-2 (0.14), but it struggles to answer specific questions compared to the improved version that has the best ROUGE-1 (0.59) and ROUGE-L (0.54). The region-focused Alpha-CLIP combined with LLaVA-Med achieved the best performance on both the ROUGE-2 (0.40) and BLEU (0.28), while still exhibiting accurate and context-sensitive reasoning, which points to the importance of task-specific adaptation in medical VQA. The combined results suggest that concentrating attention on specific areas within images, along with instruction-based reasoning, is essential for medical AI systems that manage diverse types of data. This approach facilitates the comparison of methods and provides a detailed strategy for improving specialized diagnostic medical AI assistants.

The demand for efficient and accurate Clinical Decision Support Systems (CDSS) is growing rapidly, driven by the escalating volume of medical data. While Large Language Models (LLMs) offer significant potential, their direct application in healthcare is limited by issues like hallucinations and lack of domain-specific knowledge. Retrieval-Augmented Generation (RAG) addresses these challenges by grounding LLMs with external knowledge, and recent lightweight RAG-based CDSS have shown promise. Building on this, we propose Enhanced Clinical RAG-LLM (ECRAG-LLM), a novel system designed to elevate performance in complex clinical scenarios. ECRAG-LLM utilizes a robust yet lightweight Mistral-based LLM, integrated with a multi-stage instruction tuning strategy that first adapts to general medical knowledge and then reinforces context-aware and causal reasoning using a custom dataset of structured clinical cases. We employ BioSimCSE for domain-specific embeddings and introduce an enhanced RAG architecture featuring hybrid retrieval, cross-encoder-based contextual re-ranking, and context summarization to optimize retrieved information. Extensive experiments on medical benchmarks demonstrate that ECRAG-LLM consistently outperforms baseline lightweight fine-tuned LLMs, achieving significant improvements in diagnostic accuracy, treatment appropriateness, and explanatory quality, particularly in tasks requiring deep clinical reasoning. An ablation study confirms the synergistic contributions of our innovations, and an error analysis highlights a substantial reduction in critical errors, positioning ECRAG-LLM as a more reliable and intelligent solution for resource-constrained clinical environments.

Many languages lack the vast datasets needed to develop high-performing language models, resulting in significant disparities in performance across various linguistic tasks. The novel recursive instruction tuning approach proposed in this research addresses this challenge through an iterative feedback mechanism that enhances task-specific performance for lowresource languages. By applying recursive instructions to Mistral, an open-source language model, the study demonstrates substantial improvements in translation accuracy, syntactic parsing, and other complex linguistic tasks where traditional methods often falter. The recursive framework's ability to iteratively refine instructions based on model output is shown to be highly effective, enabling more robust learning even in data-constrained environments. Through rigorous experimentation, the findings reveal that this technique leads to more adaptable and fluent models, particularly in handling the diverse grammatical and syntactic constructs found in low-resource languages. The research demonstrates the practical benefits of recursive instruction tuning, highlighting its scalability and efficacy across a variety of linguistic tasks and language groups.