102 articles
Helpful and harmless assistant
A helpful and harmless assistant is a language model trained to be useful while avoiding harmful outputs, the core of Anthropic's 2021 HHH alignment framework.
HelpSteer
HelpSteer is a family of open, human-annotated helpfulness datasets released by NVIDIA, spanning three releases from November 2023 to March 2025 under the CC-BY-4.0 license.
IA3 (machine learning)
IA3, short for Infused Adapter by Inhibiting and Amplifying Inner Activations, is a parameter-efficient fine-tuning method introduced in 2022 in which learned vectors rescale a transformer's internal activations.
Identity preference optimization
Identity preference optimization (IPO) is a preference-optimization loss for language-model alignment, introduced at Google DeepMind in 2023, that modifies DPO by targeting a fixed margin to reduce overfitting.
Infinity-Instruct
Infinity-Instruct is a large-scale instruction-tuning dataset from the Beijing Academy of Artificial Intelligence, built on the argument that coverage and depth of instructions matter more than raw sample count.
InstructGPT
InstructGPT is a family of GPT-3 models fine-tuned by OpenAI to follow written instructions using reinforcement learning from human feedback, released in January 2022; a 1.3B version was preferred over the 175B base GPT-3.
Instruction tuning
Instruction tuning, also called supervised fine-tuning (SFT), is the supervised training of a pretrained language model on instruction-output pairs so it responds to instructions instead of merely continuing text.
Kahneman–Tversky optimization
Kahneman–Tversky optimization (KTO) is a preference-optimization method for aligning large language models, introduced by Kawin Ethayarajh and colleagues in February 2024 as an alternative to DPO.
KL regularization in RLHF
KL regularization in RLHF is a penalty term that keeps a fine-tuned language model close to a frozen reference policy during reinforcement learning from human feedback.
Knowledge distillation (for LLMs)
Knowledge distillation for large language models is a training technique that transfers a large teacher model's behavior into a smaller student, introduced by Hinton, Vinyals, and Dean in 2015.
LIMA (dataset)
LIMA is a supervised fine-tuning dataset of exactly 1,000 curated prompts and responses, released in May 2023, whose 65B LLaMA model matched GPT-4 in 43% of comparisons.
LoRA adapters for generative media
LoRA (Low-Rank Adaptation) is a fine-tuning method, introduced by Microsoft in 2021, that freezes a pretrained model's weights and trains small low-rank update matrices instead.
Low-rank adaptation
Low-rank adaptation (LoRA) is a parameter-efficient fine-tuning technique, proposed by Edward J. Hu and colleagues at Microsoft in 2021, that freezes pre-trained model weights and trains small low-rank matrices.
Math-Shepherd
Math-Shepherd is a process reward model for mathematical reasoning in large language models, introduced in December 2023, that labels each reasoning step by whether model completions from it reach the correct answer.
Mergekit
MergeKit is an open-source toolkit for model merging, combining fine-tuned model weights into a single multitask model by arithmetic alone, created by Charles Goddard in 2023.
MiniLLM
MiniLLM is a knowledge-distillation method for generative large language models, introduced in 2023 by researchers at Tsinghua University and Microsoft Research, that replaces forward with reverse Kullback–Leibler divergence.
Model merging
Model merging is a post-training technique that combines the weight tensors of several fine-tuned models sharing the same base model into a single model, blending their capabilities without further training.
Model soups
Model soups is a machine learning technique introduced by Google Research in 2022 that averages the weights of several fine-tuned models, improving accuracy with no added inference cost.
Multi-stage frontier post-training pipelines
A multi-stage frontier post-training pipeline is the ordered sequence of training stages, typically six to ten, that turns a pretrained large language model into an assistant.
Nash learning from human feedback
Nash learning from human feedback (NLHF) is a method for fine-tuning language models from pairwise human preferences by seeking a Nash equilibrium instead of maximizing a reward model.
Odds ratio preference optimization
Odds ratio preference optimization (ORPO) is a single-stage preference alignment method for large language models that merges fine-tuning and preference optimization without a reference model, introduced at KAIST in 2024.
OmegaPRM
OmegaPRM is a divide-and-conquer Monte Carlo tree search algorithm introduced by Google DeepMind researchers in June 2024 to label the first error in a chain of thought for training process reward models.
On-policy distillation
On-policy distillation is a post-training method for language models in which the student generates its own rollouts and a stronger teacher grades every token, combining reinforcement learning's on-policy training with dense distillation supervision.
Online vs. offline preference learning
Online vs. offline preference learning is the distinction in AI post-training between updating on fresh rollouts from the current policy, like PPO-based RLHF, and fixed preference datasets, like DPO.
Open post-training reproduction efforts (Open-R1)
Open-R1 is a Hugging Face project launched in January 2025 to reconstruct the missing training data and pipeline behind DeepSeek-R1, whose weights were open but whose recipe was not.
OpenAssistant Conversations (OASST)
OpenAssistant Conversations (OASST) is a multilingual, human-annotated dataset of assistant conversations released by LAION's OpenAssistant project in 2023, written and rated by over 13,500 volunteers.
OpenOrca
OpenOrca is an open-source instruction-tuning dataset for large language models, built by augmenting FLAN Collection prompts with about 1 million GPT-4 and 3.2 million GPT-3.5 responses.
Outcome reward model
An outcome reward model (ORM) is a reward model that judges a language model's reasoning solely by whether its final answer is correct, ignoring intermediate steps.
P-tuning
P-tuning is a parameter-efficient fine-tuning method for pretrained language models, introduced by Tsinghua University researchers in 2021, where trainable continuous prompt embeddings are optimized with a prompt encoder while the model stays frozen.
PEFT
PEFT (parameter-efficient fine-tuning) adapts large pretrained models to new tasks by training only a small fraction of added or reparametrized parameters, achieving performance comparable to full fine-tuning.