Large Language Models
142 views
DPO
Quick Definition
Alignment method bypassing reward model for preference learning
Full Definition
Direct Preference Optimization, a simplified alignment method that bypasses explicit reward model training.
Examples
LLM alignment, preference learning, simpler RLHF alternative
Related Terms
rlhf
alignment
reward-model
More Large Language Models Terms
Model Merging
Combining fine-tuned LLMs without additional training
KV Cache
Memory optimization storing key-values for faster generation
Beam Search
Search algorithm exploring multiple output sequences
Nucleus Sampling
Generation using cumulative probability threshold for token selection
Flash Attention
Memory-efficient attention using GPU SRAM block computation
System Prompt
Special prompt defining AI assistant behavior and constraints