Luna Evaluation Models

AI model · Safety / moderation · Text → Tables / structured data by Galileo

Purpose-built small language models for AI agent evaluation and safety.

Purpose-built small language models for evaluating and guardrailing AI agent interactions. They affordably score 100% of production traffic for correctness and safety to power runtime guardrails.

Model specs

The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.

Languages
English

Category: Generative AI

Model Type
Small Language Model (SLM)
Primary Function
Evaluation and Guardrailing
Monitoring Coverage
100% of production traffic
Scoring Metrics
Correctness, safety, and tool use
  • Purpose-built small language models (SLMs) for AI evaluation
  • Scores correctness, safety, and tool use for every interaction
  • Affordably monitors 100% of production AI traffic
  • Powers runtime guardrails to stop unsafe agent actions

Overview

Luna Evaluation Models are Splunk's purpose-built small language models (SLMs) designed specifically to evaluate and guardrail AI agent interactions. These models solve the problem of high monitoring costs by providing a smaller, more efficient architecture that can score 100% of production traffic for correctness, safety, and tool use. Instead of sampling data, businesses can use Luna to monitor every interaction affordably compared to using larger frontier models. The evaluations generated by Luna models serve as the foundation for runtime guardrails, which proactively stop unsafe agent actions before they are executed. This offering is part of Splunk's broader AI and agentic operations suite, helping organizations govern AI at scale.

Evaluation and Guardrails

  • Runtime Guardrails: Luna models provide the data necessary to implement runtime guardrails that stop unsafe agent actions.
  • Affordable Monitoring: The architecture is designed to watch 100% of AI traffic affordably compared to larger frontier models.
  • Full Production Coverage: The models evaluate every interaction in production rather than relying on a sample.

Use cases

  • Security teams use Luna models to score the safety of every AI agent interaction in production to prevent data leakage.
  • Platform owners monitor 100% of AI traffic affordably to identify hallucinations without the cost of frontier models.
  • Developers implement runtime guardrails based on Luna evaluations to stop unauthorized agent tool usage in real time.

Specifications

Model Type
Small Language Model (SLM)
Primary Function
Evaluation and Guardrailing
Monitoring Coverage
100% of production traffic
Scoring Metrics
Correctness, safety, and tool use

How it works

  1. Deployment: The models are deployed within the Splunk platform to monitor AI agent and model interactions.
  2. Real-time Evaluation: Luna models analyze every interaction in production, scoring them for safety and correctness.
  3. Guardrail Implementation: Based on the trust scores from Luna, runtime guardrails are applied to prevent unauthorized or unsafe actions.

FAQ

What are Luna Evaluation Models?

Luna Evaluation Models are purpose-built small language models designed to evaluate and guardrail AI agent interactions.

How do they help control costs?

By using a smaller architecture, they allow businesses to monitor 100% of AI traffic affordably without the high costs of frontier models.

How do they improve AI safety?

Evaluations from Luna models are used to build runtime guardrails that stop unsafe agent actions before tools run.

Added by AntSpace from public sources (splunk.com).

More AI models by Galileo