Skip to main content

Trustworthy Language Model (TLM)

Even today’s best Large Language Models (LLMs) still produce “hallucinations”, occasional incorrect answers that can undermine your business. To help you achieve reliable AI, Cleanlab’s Trustworthy Language Model scores the trustworthiness of responses from any LLM in real-time. This lets you know which outputs are reliable and which ones need extra scrutiny.

Benchmarks reveal that TLM can reduce the rate of incorrect answers: of GPT-4o by 27%, of o1 by 20%, and of Claude 3.5 Sonnet by 20%. For RAG applications, TLM detects incorrect answers with 3x greater precision than alternatives including: RAGAS, LLM-as-judge, G-Eval, DeepEval, HHEM, Lynx, Prometheus-2, or LogProbs.

Get started using the TLM API via our quickstart tutorial.

Key Features​

  • Trustworthiness Scores: TLM provides state-of-the-art trustworthiness scores for any LLM application (Q&A, RAG, Classification, Summarization, Data Extraction, Structured Outputs, Data Annotation, …). You can score responses from your existing model (works for any LLM) or even those written by humans.
  • Improved LLM Responses: You can alternatively use TLM in place of your own LLM to produce higher accuracy outputs (along with trustworthiness scores). TLM can outperform every frontier model, because it is built on top of these models, refining their generation process via trustworthiness scoring.
  • Scalable Real-Time API: Designed to handle large datasets, TLM is suitable for most enterprise applications and offers flexible configurations to control costs/latency, as well as private deployment options.

Unlocking Enterprise Use Cases​

  • Retrieval-Augmented Generation: TLM automatically flags potentially incorrect responses in any RAG application. Ensure users don’t lose trust in your Q&A system.
  • Chatbots: TLM informs you which LLM outputs you can directly rely on and which LLM outputs you might escalate for human review. With standard LLM APIs, you do not know which outputs to trust.
  • Data Labeling: TLM auto-labels data with high accuracy and reliable confidence scores. Let the LLM automatically handle 99% of cases where it is trustworthy, and manually handle the remaining 1%.
  • Data Extraction and Structured Outputs: TLM tells you which data auto-extracted from documents, databases, transcripts is worth double checking to ensure accuracy. Transform raw unstructured data into structured data, with less errors and 90% less time reviewing results.
  • Evals: Use TLM to quickly find bad LLM responses in your logs, the key step in determining how to improve your AI application. TLM outperforms LLM-as-judge, and can account for custom evaluation criteria and human feedback.

TLM is effective in any LLM application where reliability is key, including many more use-cases: summarization, classification, function calling, code generation, …

Start using TLM in under a minute via our quickstart tutorial.