> ## Documentation Index
> Fetch the complete documentation index at: https://docs.trymonte.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Metrics

> Training and eval metrics

Every run records metrics. A training run records the trainer's metrics at each step, and an eval run records its score and checks on the responses. The Console charts both, and `monte run show` prints them in the terminal.

## Training Metrics

The trainer writes its metrics at every step, grouped into families by prefix. The exact set depends on the method and the trainer.

| Family | Description | Examples |
| - | - | - |
| `train/*` | The model under training | `train/reward`, `train/loss`, `train/kl_penalty`, `train/grad_norm`, `train/entropy`, `train/truncation_rate` |
| `performance/*` | Throughput on the GPU instance | |
| `timing/*` | Time in each phase of a step | |

## Eval Metrics

| Metric | Description |
| - | - |
| Score and interval | The mean grader score over tasks, with its 95% half-width |
| `mean_output_tokens` | The mean response length in tokens |
| `parse_fail_rate` | The share of responses where the grader found no answer |
| `truncations` | The share of responses that stopped early |
| `repetition` | A measure of repeated text in the responses |

The checks explain a score change. For example, a rise in `parse_fail_rate` with a fall in score shows a format problem, not a quality problem.

## View Metrics

In the Console, click **Training**, select the experiment and the run, then click **Metrics**. Each series has a chart that marks the checkpoint steps. Click the star on a chart to pin it for every run.

In the terminal:

```bash theme={null}
monte run show support-triage/train-07
monte run show support-triage/train-07 --json --metrics train/reward,train/kl_penalty
monte run logs support-triage/train-07 -f
```

`--json` with `--metrics` prints each series as a list of `{"step": n, "value": x}` points. `monte run logs` prints the run's events as JSON lines.

## Compare Runs

`monte compare` takes two eval runs on the same eval. It shows the [paired difference](https://docs.trymonte.ai/evaluation/scoring-models.md#read-the-result) and each task that one run passed and the other failed. The first run is the reference.

```bash theme={null}
monte compare support-triage/eval@0 support-triage/eval-03
```

```text theme={null}
difference  +0.10 ± 0.06 paired   30 fixed / 10 broken / 160 same
```

Both runs must be on the same eval, in the same experiment, and from the same base model, and neither can be a smoke run. Otherwise the command refuses and names the condition.

To compare in the Console:

1. Click **Evaluations** and select the experiment.
2. Select two eval runs on the same eval.
3. Click **Compare**.

The view lists each task as **Fixed**, **Broken**, or **Same**. Click a row to open the two [traces](https://docs.trymonte.ai/observability/traces.md) side by side.


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