> ## Documentation Index
> Fetch the complete documentation index at: https://mirage-eval.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Metrics

> The four hallucination metrics Mirage implements.

Mirage implements four object hallucination metrics. Each keeps its paper's questions and scoring, and adds a named modification where sparse labels break it. Every run reports the raw paper score next to the modified score.

| Metric | What the model does | Headline score |
| - | - | - |
| [POPE](https://mirage-eval.com/metrics/pope.md) | Answers yes or no: is this object in the image? | Mod-POPE balanced accuracy |
| [ROPE](https://mirage-eval.com/metrics/rope.md) | Names the object inside a marked box | Mod-ROPE accuracy |
| [CHAIR](https://mirage-eval.com/metrics/chair.md) | Writes a caption | Mod-CHAIR F0.5 |
| [THRONE](https://mirage-eval.com/metrics/throne.md) | Writes a caption | THRONE class-wise F0.5 |

The data behind each change is on [Why the metrics are modified](https://mirage-eval.com/metrics/why.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.