LinkTransformer
linktransformer.github.io
What LinkTransformer is
LinkTransformer is a Python package for record linkage and other operations on tabular data. It supports semantic matching, candidate retrieval, row transformation, clustering, and text classification, with dataframe workflows that include deduplication and multilingual merges. For classification, it can use Hugging Face or OpenAI models, and users can train a custom classifier. The package works with models on Hugging Face Hub and OpenAI embedding models; the repository also documents Gemini API keys, API embedding models, and local SBERT models. Users can train linkage models on their own data or fine-tune pretrained Hugging Face models, then share trained models through Hugging Face Hub. The repository gives `pip install linktransformer` as the installation command, and Python 3.11 or later is required. Tutorials cover record linkage, model training, and text classification, alongside a Hugging Face demo app. LinkTransformer is free and open source under the MIT License. Its website lists additional capabilities, including hard-negative mining and GPU and RAPIDS integrations, as coming soon.
Who it suits
LinkTransformer suits Python users working with duplicate records, clustering, merges, or text classification in tabular data. It may also fit teams that want to train and share their own linkage models.
Good to know, and what to watch
- Supports dataframe deduplication, clustering, and merges
- Uses Hugging Face and OpenAI models
- Allows custom classifier training
- Free and MIT-licensed
- Includes tutorials and a demo app
- Requires Python 3.11 or later
- Some listed capabilities are still coming soon
The verdict
LinkTransformer brings record linkage, dataframe operations, and classification into a Python package, with options to train and share models. Check the Python requirement and the status of coming-soon capabilities against your needs.
What the maker publishes
- Coming soon
- The site lists FAISS GPU, cuDF, cuML, and cuGraph integration as coming soon. source
- Custom training
- Users can train record linkage models on their own data or fine-tune pretrained Hugging Face models. source
- Custom training
- Users can train record linkage models on their own data or fine-tune pretrained Hugging Face models. source
- Data operations
- It supports dataframe deduplication, clustering, and merges. source
- Installation
- The repository gives `pip install linktransformer` as the installation command. source
- Installation
- The site gives `pip install linktransformer` as the installation command. source
- Integrations
- The repository documents OpenAI and Gemini API keys, API embedding models, and local SBERT models. source
- Integrations
- The site describes integrations with Hugging Face Hub, OpenAI models, and Hugging Face models. source
- Learning resources
- The repository links tutorials for record linkage, training a model, and text classification, plus a Hugging Face demo app. source
- License
- The repository states that the project is licensed under the MIT License. source
- Model sharing
- Trained models can be uploaded to or downloaded from Hugging Face Hub, where LinkTransformer models receive a discoverability tag. source
- Model sharing
- The package integrates with Hugging Face Hub for uploading and downloading models. source
- Model support
- The website says LinkTransformer supports models on Hugging Face Hub and OpenAI embedding models. source
- Model support
- It supports models from Hugging Face Hub and OpenAI embedding models. source
- Open source
- The project repository states that LinkTransformer is licensed under the MIT License. source
- Product
- LinkTransformer is a Python package for semantic record linkage, candidate retrieval, row transformation, clustering, and text classification over tabular data. source
- Purpose
- LinkTransformer is a Python package for record linkage using transformer models. source
- Requirements
- The package listing states that Python 3.11 or later is required. source
- Support
- The site invites users to contribute through the repository and raise issues with feature requests. source
- Support and contribution
- The website invites users to collaborate, and the repository says issues and pull requests are welcome. source
- Text classification
- The package can classify dataframe rows using Hugging Face or OpenAI models and supports custom classifier training. source
- Text classification
- It can classify dataframe rows with OpenAI and Hugging Face models, including binary and multiclass classification. source
- Upcoming capabilities
- The website lists hard-negative mining, GPU and RAPIDS integrations, convenience wrappers, and vision or multimodal support as coming soon. source
- Use cases
- The maker’s website describes dataframe operations including deduplication, clustering, and multilingual merges. source
Checked against the maker's pages on 9 Oct 2026.