LatitudeGames has launched Equinox-31B, a new AI model finetuned from Gemma 31B, designed to offer a balanced experience between dark adventures and slice-of-life storytelling. The model is versatile and available for use on the AIDungeon platform, requiring a subscription.
The Piggyback Hypothesis explains how chat-template tokens can cause emergent misalignment in LLMs by generalizing finetuned behavior to out-of-domain queries. Token-Regularized Finetuning (TReFT) is proposed to mitigate this issue, preserving in-domain learning while reducing misalignment across models and datasets.
The user expresses great satisfaction with local AI, emphasizing the freedom to finetune models without censorship or data harvesting, ensuring privacy for personal discussions. They are eternally grateful to developers of open-weight models and tools like llama.cpp for these benefits.
This paper introduces a perplexity-based method to reveal finetuning objectives of large language models, particularly those exhibiting "model organism" behaviors. This method leverages models' tendency to overgeneralize, generating and ranking completions to identify the finetuning goals without prior assumptions.
This paper introduces Foundation Preserving LoRA (FoLoRA), an optimization framework that addresses the degradation of nontarget capabilities during finetuning of foundation models. It uses a generalized Rayleigh quotient to balance task utility and forgetting penalty, guiding updates to preserve pretraining knowledge.
"The Developer's Guide to Finetuning LLMs" provides a practical guide for developers and retail leaders on when, why, and how to finetune LLMs. It covers decision frameworks, costs, data preparation, and evaluation strategies, comparing finetuning with prompt engineering and RAG for brand-specific tasks.
Sparse Memory Finetuning (SMF) addresses catastrophic forgetting in pretrained language models by updating only a small subset of memory rows. Experiments show SMF improves performance on a medical exam task while substantially mitigating forgetting compared to LoRA and full finetuning.
This content discusses the training and finetuning of multimodal embedding and reranker models, leveraging the Sentence Transformers library to optimize their performance.