thunlp/OpenDelta

Health v2 · maintenance only

Adoption not part of the score

1046 stars · 83 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

OpenDelta is a Python library for parameter-efficient tuning (delta tuning) of pretrained language models, letting users attach small trainable modules like adapters, LoRA, or prefix-tuning while freezing the rest. It plugs into PyTorch and Hugging Face Transformers models with flexible, name-based module addressing.

Use cases

• fine-tune large language models with LoRA instead of full fine-tuning
• add adapters to a pretrained transformer while freezing base weights
• implement prefix tuning or soft prompt tuning on BART or T5
• serve multiple tasks from one frozen PLM with small delta modules
• reduce GPU memory and storage costs when fine-tuning pretrained models
• experiment with parameter-efficient learning methods on Hugging Face models

When to choose

• you want plug-and-play parameter-efficient tuning on PyTorch/Transformers models
• you need to swap between adapter, LoRA, prefix-tuning, and other delta methods
• you want space-saving multitask serving with a shared frozen backbone

When to avoid

• you need full fine-tuning of all model parameters
• you work outside PyTorch or with non-Transformers model architectures
• you need actively maintained support for the latest LLM architectures

Facets

library · maturity maintenance

machine-learning deep-learning nlp machine-learning deep-learning python parameter-efficient-tuning delta-tuning lora adapters prefix-tuning pytorch transformers prompt-tuning natural-language-processing

Member repositories

For agents

markdown · JSON · MCP: product_card(name="thunlp/OpenDelta")