A simplified alternative to RLHF that aligns LLM outputs with human preferences without training a separate reward model or using reinforcement learning. DPO directly optimizes a policy using pairs of preferred and dispreferred outputs, making it computationally cheaper and more stable than RLHF's multi-stage pipeline. Widely adopted in 2024-2025 for fine-tuning open-source models.
Start with the shortest useful explanation before going deeper.
A simplified alternative to RLHF that aligns LLM outputs with human preferences without training a separate reward model or using reinforcement learning. DPO directly optimizes a policy using pairs of preferred and dispreferred outputs, making it computationally cheaper and more stable than RLHF's multi-stage pipeline. Widely adopted in 2024-2025 for fine-tuning open-source models.
Use the quick analogy first so the term is easier to reason about when you meet it in code, docs, or prompts.
Think of it as a piece of the context or inference stack behind agentic and LLM-powered Solana products.
Place the term inside its Solana layer so the definition is easier to reason about.
LLMs, RAG, embeddings, inference, and agent-facing primitives.
Turn the term from vocabulary into something operational for product and engineering work.
This term unlocks adjacent concepts quickly, so it works best when you treat it as a junction instead of an isolated definition.
Use this compact block when you want to give an agent or assistant grounded context without dumping the entire page.
DPO (Direct Preference Optimization) (dpo) Category: AI / ML Definition: A simplified alternative to RLHF that aligns LLM outputs with human preferences without training a separate reward model or using reinforcement learning. DPO directly optimizes a policy using pairs of preferred and dispreferred outputs, making it computationally cheaper and more stable than RLHF's multi-stage pipeline. Widely adopted in 2024-2025 for fine-tuning open-source models. Aliases: Direct Preference Optimization Related: RLHF (Reinforcement Learning from Human Feedback), Fine-Tuning, Training (ML)
Use glossary context, relationships, mental models, and builder paths to get structured answers instead of generic chat output.
Optional: paste Anchor, Solana, or Rust code so the Copilot can map primitives back to glossary terms.
The Copilot will answer using the current term, related concepts, mental models, and the surrounding glossary graph.
These branches show which concepts this term touches directly and what sits one layer beyond them.
A training technique that aligns LLM outputs with human preferences. Process: (1) train a reward model from human comparisons of outputs, (2) use reinforcement learning (PPO) to optimize the LLM against the reward model. RLHF makes models more helpful, harmless, and honest. Used by Claude, ChatGPT, and other assistants. Alternatives include DPO (Direct Preference Optimization) and Constitutional AI.
The process of further training a pre-trained model on a specialized dataset to improve performance on specific tasks. Fine-tuning adapts a foundation model's weights using domain-specific data (e.g., Solana documentation, smart contract code). Techniques include full fine-tuning, LoRA (Low-Rank Adaptation), and QLoRA. Fine-tuned models can outperform general models on narrow tasks.
The process of optimizing a model's parameters by exposing it to data and adjusting weights to minimize a loss function. Pre-training on large datasets creates foundation models. Training LLMs requires massive compute (thousands of GPUs, weeks/months). Training data quality, diversity, and size directly impact model capabilities. Distinguished from fine-tuning (smaller scale, specific domain).
These are the next concepts worth opening if you want this term to make more sense inside a real Solana workflow.
Glossary entries become useful when they are connected. These links are the shortest path to adjacent ideas.
A training technique that aligns LLM outputs with human preferences. Process: (1) train a reward model from human comparisons of outputs, (2) use reinforcement learning (PPO) to optimize the LLM against the reward model. RLHF makes models more helpful, harmless, and honest. Used by Claude, ChatGPT, and other assistants. Alternatives include DPO (Direct Preference Optimization) and Constitutional AI.
The process of further training a pre-trained model on a specialized dataset to improve performance on specific tasks. Fine-tuning adapts a foundation model's weights using domain-specific data (e.g., Solana documentation, smart contract code). Techniques include full fine-tuning, LoRA (Low-Rank Adaptation), and QLoRA. Fine-tuned models can outperform general models on narrow tasks.
The process of optimizing a model's parameters by exposing it to data and adjusting weights to minimize a loss function. Pre-training on large datasets creates foundation models. Training LLMs requires massive compute (thousands of GPUs, weeks/months). Training data quality, diversity, and size directly impact model capabilities. Distinguished from fine-tuning (smaller scale, specific domain).
These entries live beside the current term and help the page feel like part of a larger knowledge graph instead of a dead end.
A neural network trained on vast text corpora to understand and generate human language. LLMs (GPT-4, Claude, Llama, Gemini) use transformer architectures with billions of parameters. They power chatbots, code generation, summarization, and reasoning tasks. In blockchain development, LLMs assist with smart contract writing, audit review, documentation, and code explanation.
The neural network architecture underlying modern LLMs, introduced in 'Attention Is All You Need' (2017). Transformers use self-attention mechanisms to process input sequences in parallel (unlike recurrent networks). Key components: multi-head attention, positional encoding, feedforward layers, and layer normalization. Variants include encoder-only (BERT), decoder-only (GPT), and encoder-decoder (T5).
A neural network component that allows models to weigh the relevance of different parts of the input when producing output. Self-attention computes query-key-value dot products across all positions, enabling each token to 'attend' to every other token. Multi-head attention runs multiple attention functions in parallel. Attention is O(n²) in sequence length, driving context window research.
A large AI model trained on broad data that can be adapted for many downstream tasks. Foundation models (GPT-4, Claude, Llama 3, Gemini) are pre-trained on internet-scale text/code and can be fine-tuned, prompted, or used via APIs for specific applications. The term emphasizes that one base model serves as the foundation for diverse use cases rather than training task-specific models.