This paper analyzes the interpretive behaviors of LLMs for automated code compliance using perturbation-based attribution analysis, comparing different fine-tuning strategies and model scales. Results show full fine-tuning yields more focused attribution patterns, and larger models prioritize specific textual elements like numerical constraints.
This content provides a guide on how to fine-tune the Nemotron 3.5 Automatic Speech Recognition (ASR) model. It aims to help users adapt the model for specific languages, domains, or accents, optimizing its performance.
This paper introduces TeamTR, a trust-region framework for fine-tuning multi-agent LLM systems, addressing structural failures in sequential fine-tuning. It proves that stale-occupancy evaluation incurs a quadratic penalty with the number of agents and improves performance by 7.1% on average.
LoRA (Low-Rank Adaptation) is introduced as the leading PEFT method, enabling efficient adaptation of massive LLMs like Llama 3 without requiring extensive hardware resources. The post promises to delve into LoRA's mathematical intuition, the concept of "intrinsic dimension," and its game-changing impact for AI engineers.
Large language models often hallucinate facts, a problem exacerbated by supervised fine-tuning (SFT) which degrades pre-trained knowledge. This research proposes a self-distillation SFT method, inspired by continual learning, to mitigate hallucinations by regularizing output-distribution drift while effectively acquiring new factual information.
This work introduces Group Fine-Tuning (GFT), a unified post-training framework for large language models. It addresses intrinsic limitations of supervised fine-tuning (SFT), such as single-path dependency and entropy collapse, through Group Advantage Learning and Dynamic Coefficient Rectification.
TalkLoRA propõe um framework MoELoRA que aborda a instabilidade de roteamento e a dominância de especialistas em métodos existentes, permitindo a comunicação entre especialistas antes do roteamento. Isso é feito através de um Módulo de Conversação leve, que facilita a troca de informações, gerando um sinal de roteamento mais robusto para Large Language Models (LLMs).
This research discovers that LoRA fine-tuning leads to 'un-learning' on contested examples, where high annotator disagreement correlates with increased loss during training. This pattern is distinct from full fine-tuning and consistently observed across multiple encoder and decoder-only models and datasets.
HELLoRA proposes a novel method for fine-tuning Mixture-of-Experts (MoE) models by applying Low-Rank Adaptation (LoRA) modules only to the most frequently activated experts at each layer. This technique significantly reduces trainable parameters and improves downstream performance, attributing its success to structured regularization that maintains expert specialization.
This article analyzes the choice between RAG, fine-tuning, and AI agents for LLM projects, suggesting that a combination is often needed. It provides a practical guide on which architecture to prioritize based on project needs such as data source, actions, and budget.
This paper presents QU-NLP's multi-stage QLoRA fine-tuning strategy for Arabic Islamic inheritance reasoning using Qwen3-4B. The model achieved a 90% MIR-E score, demonstrating competitive performance with minimal computational resources.
This content details the fine-tuning of a clinical AI model, MedQA, on the AMD ROCm platform. It highlights the ability to perform this task without requiring CUDA, offering a significant alternative for AI development.
This article highlights how user 'thumbs down' feedback provides invaluable real-world training data for AI systems, often overlooked by teams. It challenges the misconception that AI fine-tuning is always expensive and complex, proposing a simplified feedback loop suitable for product development.
This article details the process of fine-tuning OpenCLIP ViT-B/32 for architectural styles, achieving a +26 percentage point increase in accuracy. The author focuses on the critical decisions made before and after the training loop that were responsible for this significant result, rather than the training loop optimization itself.
This post explores the art and science of hyperparameter optimization on Amazon Nova Forge, detailing how to balance improving domain-specific performance without degrading a model's general capabilities. It covers customization strategies, configuring training parameters like learning rate and batch size, and avoiding common mistakes that lead to wasted training runs.
Recent work demonstrates that post-training tricks can significantly cut LLM cost and memory footprint without losing ability. These include aligning synthetic data with a student's style and utilizing key-value (KV) cache optimizations, achieving substantial savings without typical performance drops.
This paper details an attempt to distill behavioral dispositions into small language models (0.6B-2.3B parameters) through a distillation pipeline. Initial reported gains were later falsified due to evaluation artifacts, resulting in a negative outcome for the core hypothesis and leading to three subsequent arcs of investigation.
This research challenges the assumption that Parameter-Efficient Fine-Tuning (PEFT) equates to memory efficiency for on-device LLMs, showing existing methods can still lead to out-of-memory errors. It introduces LARS (Low-memory Activation-Rank Subspace), a novel framework that decouples memory consumption from sequence length by constraining the activation subspace, achieving an average 33.54% memory footprint reduction.
This research investigates the training-time mechanisms of refusal in safety-aligned language models, specifically comparing supervised fine-tuning with R2D2-style dynamic adversarial fine-tuning. Findings show R2D2 initially achieves strong refusal on HarmBench but then partially reopens, while SFT remains consistently less robust.
Este estudo avalia metodologias de Large Language Models (LLM) – Fine-Tuning, RAG e uma abordagem Híbrida – para construir uma base de conhecimento de Análise de Causa Raiz (RCA) a partir de tickets de suporte. Os experimentos com um conjunto de dados industrial real demonstram que a base de conhecimento gerada acelera as tarefas de RCA e melhora a resiliência da rede.