Metrics

Foundry supports displaying metrics from your experiment as live plots in the web UI. To log metrics, write JSON lines to foundry/metrics.jsonl:

{"ts": 1709012345.123, "name": "train/loss", "value": 0.342, "step": 100}
{"ts": 1709012345.123, "name": "train/lr", "value": 0.001, "step": 100}
{"ts": 1709012400.456, "name": "eval/accuracy", "value": 0.89, "step": 100}

Project Root

The project root is the directory containing your .foundryconfig file. Foundry sets this as the working directory when running your experiment, so relative paths like foundry/metrics.jsonl resolve from here automatically — both on remote GPU instances and locally.

my-experiment/              ← project root (contains .foundryconfig)
├── .foundryconfig
├── train.py
└── foundry/
    ├── metrics.jsonl       ← write metrics here
    ├── outputs/
    └── checkpoints/

Schema

Each line in metrics.jsonl is a JSON object with the following fields:

Field	Required	Type	Description
name	yes	string	Metric name, e.g. train/loss
value	yes	number	Numeric value
step	no	number	Training step or iteration
ts	no	number	Unix timestamp (seconds)

You can include any additional fields you like (e.g. epoch, split, run). Foundry stores them as metadata and makes them available for filtering in the UI.

Naming Convention

Use / to namespace your metrics. The web UI groups metrics by the prefix before the first slash:

{"ts": 1709012345.123, "name": "train/loss", "value": 0.34, "step": 100}
{"ts": 1709012345.123, "name": "train/lr", "value": 0.001, "step": 100}
{"ts": 1709012400.456, "name": "eval/loss", "value": 0.52, "step": 100}
{"ts": 1709012400.456, "name": "eval/accuracy", "value": 0.89, "step": 100}
{"ts": 1709012400.456, "name": "gpu/memory_gb", "value": 22.4, "step": 100}

This would create three panels in the UI: train, eval, and gpu.

Logging Metrics from Python

Here’s a minimal implementation you can drop into your training script:

import json
import time
import os

os.makedirs("foundry", exist_ok=True)

def log_metric(name, value, step=None, **kwargs):
    m = {"ts": time.time(), "name": name, "value": value}
    if step is not None:
        m["step"] = step
    m.update(kwargs)
    with open("foundry/metrics.jsonl", "a") as f:
        f.write(json.dumps(m) + "\n")

Then call it from your training loop:

for step, batch in enumerate(dataloader):
    loss = train_step(batch)
    log_metric("train/loss", loss.item(), step=step)

    if step % eval_interval == 0:
        acc = evaluate(model, val_set)
        log_metric("eval/accuracy", acc, step=step, split="val")