Skip to main content
Grounded fact checking asks whether a document supports a claim. You give the check a document, called the context, and one sentence, called the claim. The answer is yes if the context supports the claim and no if it does not. In retrieval-augmented generation (RAG), a model answers from documents it looked up, so a claim that those documents do not support is a hallucination. Bespoke-MiniCheck-7B is a 7B model for this check. It takes a context and a claim and returns the probability that the context supports the claim. Bespoke Labs trained it from InternLM 2.5 7B Chat, based on the method in the MiniCheck paper (EMNLP 2024).

​
How to use the model well

  • Check one sentence at a time. Split a longer claim into sentences first.
  • The context can be up to 32K tokens. Split a longer document into chunks.
  • A support probability of 0.5 or more usually counts as supported. Change this threshold to fit your use case.

​
Benchmark

On the LLM-AggreFact leaderboard, Bespoke-MiniCheck-7B has an average balanced accuracy of 77.4%. As of September 2026, this is the highest average of the 39 models on the leaderboard. Most of those models were released in 2024 or earlier.

​
License

The model weights are under the CC BY-NC 4.0 license, which does not allow commercial use. For a commercial license, email company@bespokelabs.ai.

​
Ways to run MiniCheck

Self-hosting

Run the model on your own GPU with the MiniCheck library or vLLM.

Ollama

Run the model on your own computer with Ollama.

Hosted API

Call the model through the Bespoke Labs API.