Guide
Thrad Open-Sources Two Conversation Classifier Models on Hugging Face
Thrad has published two open-source models on its Hugging Face page, as part of an ongoing effort to open-source models for ads-in-AI safety.
- Published
- Aug 2026
- Reading time
- 2 min
The goal, as Thrad's Hugging Face org page puts it, is to give the community tools that make the AI ecosystem safer for advertisers and businesses, and more private for users.
Both models tackle the same underlying task: conversation intent classification — sorting a piece of conversational text into the category it belongs to.
That kind of classification is a basic building block for the wider category of work Thrad operates in, informing decisions like what a conversation is actually about, which is the foundation for keeping ad matching contextual and for identifying conversations advertising should stay out of altogether.
The two models are published together under Thrad's Conversation Classifiers collection, both released under the Apache 2.0 license — meaning anyone can download, inspect, and build on them freely.
thrad-bert-conversation-classifier
The larger of the two models, at roughly 0.1B parameters. It's built for multi-class conversation intent classification with hard labels, and is set up to run through Hugging Face's standard transformers library with a few lines of Python — load the tokenizer and model, pass in a piece of conversation text, and get back a classification. Full details on data preprocessing, training, and evaluation are published in the project's public GitHub repository. Since it was published, it has been downloaded 230 times.
Model page: huggingface.co/Thrad/thrad-bert-conversation-classifier.
thrad-distilbert-conversation-classifier
A smaller, distilled version of the same classification task, at roughly 67M parameters — about a third the size. It's tagged for knowledge distillation and ships with an ONNX export alongside the standard Safetensors weights, which makes it a more practical fit for teams that need faster or lighter-weight inference without setting up a full-size model. Same task, same license, same usage pattern via transformers.
Model page: huggingface.co/Thrad/thrad-distilbert-conversation-classifier.
Both models, along with anything else Thrad open-sources going forward, will live on Thrad's Hugging Face page.
If you're building anything that needs to understand what a conversation is about, they're free to use, inspect, and extend.
