Lucie Charlotte Magister
@charlottemagister.bsky.social
PhD student @ University of Cambridge, focusing on Explainability and Interpretability for GNNs
4/🧵Find more details and our results at: arxiv.org/abs/2411.13405
On the Way to LLM Personalization: Learning to Remember User Conversations
Large Language Models (LLMs) have quickly become an invaluable assistant for a variety of tasks. However, their effectiveness is constrained by their ability to tailor responses to human preferences a...
arxiv.org
November 21, 2024 at 6:03 PM
4/🧵Find more details and our results at: arxiv.org/abs/2411.13405
3/🧵PEFT: We use these QA pairs to finetune a LoRA adapter conversation by conversation. We find that weighting the loss on the QA tokens focuses the model on relevant content, rather than structure. Iterating on each conversation for 10 epochs gives the best results.
November 21, 2024 at 6:03 PM
3/🧵PEFT: We use these QA pairs to finetune a LoRA adapter conversation by conversation. We find that weighting the loss on the QA tokens focuses the model on relevant content, rather than structure. Iterating on each conversation for 10 epochs gives the best results.
2/🧵Data Augmentation: We up-sample conversations as positive and negative QA pairs. Positive pairs contain questions about the conversation content, while negative pairs contain questions about related topics not discussed. The negative pairs allow to draw a knowledge boundary.
November 21, 2024 at 6:03 PM
2/🧵Data Augmentation: We up-sample conversations as positive and negative QA pairs. Positive pairs contain questions about the conversation content, while negative pairs contain questions about related topics not discussed. The negative pairs allow to draw a knowledge boundary.
1/🧵PLUM is a 2 stage pipeline performing data augmentation for up-sampling conversations as QA pairs, that we use to finetune a LoRA adapter with a weighted CE loss. We perform competitively with baselines such as RAG, achieving 81.5% accuracy across 100 conversations.
November 21, 2024 at 6:03 PM
1/🧵PLUM is a 2 stage pipeline performing data augmentation for up-sampling conversations as QA pairs, that we use to finetune a LoRA adapter with a weighted CE loss. We perform competitively with baselines such as RAG, achieving 81.5% accuracy across 100 conversations.