Instruction Tuning & RLHF
Post-training methods that tune language models to follow instructions and human preferences using demonstrations, rankings, and reward models.
Core metadata
- ID: instruction_tuning_rlhf
- Era: Modern
- First known date: 2020 (exact)
- Region: Global AI research community
- Review status: source_checked
- Maturity: established
Prerequisites
- Large Language Models (large_language_models)
- Reinforcement Learning (reinforcement_learning)
- Supervised Learning Pipelines (supervised_learning_pipelines)
Dependents
Fields
Field lanes
- Artificial Intelligence & Machine Learning: Safety & Governance
Node sources
- Learning to summarize with human feedback (arXiv / OpenAI, 2020, primary_paper) • Supports: node, maturity
- Finetuned Language Models Are Zero-Shot Learners (arXiv / Google Research, 2021, primary_paper) • Supports: node, maturity
- Training language models to follow instructions with human feedback (arXiv / OpenAI, 2022, primary_paper) • Supports: node, maturity
Prerequisite edge evidence
Edge/source evidence summary:
- Prerequisite edges: 3
- Average edge confidence: 80%
- Prerequisite sources: 7
- primary_source: 3
| Prerequisite | Type | Confidence | Evidence level | Note | Sources |
|---|---|---|---|---|---|
| Large Language Models (large_language_models) | enabling | 82% | primary_source | Instruction tuning and RLHF are post-training methods applied to pretrained language models; large language models are the main modern substrate, but the methods are not a hardware-like prerequisite. |
|
| Reinforcement Learning (reinforcement_learning) | enabling | 78% | primary_source | Reinforcement learning and reward modeling underpin the RLHF part of the bundled node, while instruction tuning can also use supervised demonstrations. |
|
| Supervised Learning Pipelines (supervised_learning_pipelines) | enabling | 80% | primary_source | Instruction tuning and preference-model pipelines rely on curated demonstrations, rankings, and supervised fine-tuning workflows. |
|
This page is generated from canonical era JSON and is indexable by URL.