Cleanlab

Cleanlab's Trustworthy Language Model, known as TLM, scores the trustworthiness of responses from any large language model in real time, so an application can tell which outputs are reliably correct and which need extra scrutiny. It can score answers produced by an existing model, or even written by humans, or it can be used in place of the model itself to generate responses that come with a trustworthiness score attached. It is aimed at teams putting LLMs into production who need to catch hallucinations before they reach customers. The documentation lists coverage for retrieval augmented generation, agents, chatbots, summarization, data extraction, structured outputs, tool calls, classification, data labeling and LLM evaluations. TLM needs no training on the customer's data, so there is no dataset preparation or labeling step, and it works with any base model, including reasoning models. Quality presets, a lighter TLM Lite mode and a choice of base models let teams trade accuracy against latency and cost, and custom evaluation criteria can be added. Access is through a Python API, with private deployment options for enterprises. TLM can be tried for free: signing up for a Cleanlab account provides an API key and a grant of free tokens. After that, usage continues on a pay per token plan whose rates are shown only inside the account under Usage and Billing, not on a public page. Enterprise subscriptions with volume discounts, private deployment and additional features are quote based through the sales team. Cleanlab has been acquired by Handshake AI, and the founders now lead Handshake's AI research into training data quality. The company states that its open source data centric AI package remains available under the Apache 2.0 licence.