We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
- pizza-yogurt
Interactive charts with Vega-Lite in Elixir Phoenix
2026-04-01 ･ blog, vega-lite, elixir, phoenix, side-project, mdex, markdown
After adding diagrams1, I wished I had nicer charts2.
Interactive charts are rendered from vl fenced code blocks. The spec is validated at compile time, so any local data URLs are checked for existence too. The data lives in priv/static/data/.
Examples
Here’s some examples.
Loss curves
A multi-series line chart using Vega-Lite’s fold transform to pivot train_loss and val_loss from wide to long format.
Source
```vl
{
"$schema": "https://vega.github.io/schema/vega-lite/v5.json",
"data": { "url": "/data/vega-example/training.csv" },
"transform": [
{
"fold": ["train_loss", "val_loss"],
"as": ["split", "loss"]
}
],
"mark": { "type": "line", "point": true },
"encoding": {
"x": { "field": "epoch", "type": "quantitative", "title": "Epoch" },
"y": { "field": "loss", "type": "quantitative", "title": "Loss" },
"color": { "field": "split", "type": "nominal", "title": "Split" },
"tooltip": [
{ "field": "epoch", "type": "quantitative" },
{ "field": "split", "type": "nominal" },
{ "field": "loss", "type": "quantitative", "format": ".4f" }
]
}
}
```
Accuracy
Same pattern for accuracy, with an interactive slider to set a reference threshold line.
Source
```vl
{
"$schema": "https://vega.github.io/schema/vega-lite/v5.json",
"data": { "url": "/data/vega-example/training.csv" },
"params": [
{
"name": "threshold",
"value": 65,
"bind": {
"input": "range",
"min": 40,
"max": 80,
"step": 1,
"name": "Accuracy threshold (%)"
}
}
],
"layer": [
{
"transform": [
{ "fold": ["train_acc", "val_acc"], "as": ["split", "accuracy"] }
],
"mark": { "type": "line", "point": true },
"encoding": {
"x": { "field": "epoch", "type": "quantitative", "title": "Epoch" },
"y": { "field": "accuracy", "type": "quantitative", "title": "Accuracy (%)" },
"color": { "field": "split", "type": "nominal", "title": "Split" },
"tooltip": [
{ "field": "epoch", "type": "quantitative" },
{ "field": "split", "type": "nominal" },
{ "field": "accuracy", "type": "quantitative", "format": ".2f" }
]
}
},
{
"mark": { "type": "rule", "strokeDash": [4, 4], "color": "gray" },
"encoding": {
"y": { "datum": { "expr": "threshold" }, "type": "quantitative" }
}
}
],
"config": { "view": { "padding": { "top": 10 } } }
}
```
Learning rate schedule
The ReduceLROnPlateau scheduler halving the learning rate is visible as discrete steps.
Source
```vl
{
"$schema": "https://vega.github.io/schema/vega-lite/v5.json",
"data": { "url": "/data/vega-example/training.csv" },
"mark": { "type": "line", "point": true, "color": "steelblue" },
"encoding": {
"x": { "field": "epoch", "type": "quantitative", "title": "Epoch" },
"y": {
"field": "lr",
"type": "quantitative",
"title": "Learning rate",
"scale": { "type": "log" }
},
"tooltip": [
{ "field": "epoch", "type": "quantitative" },
{ "field": "lr", "type": "quantitative", "format": ".6f" }
]
}
}
```
-
It would’ve been nice for First shell command prediction model ↩