Fix run selection issue

[abidlabs]

Closes: #395

Repro (original issue and fix) with:

import math
import random
import time

import trackio as wandb

EPOCHS = 20
PROJECT_ID = random.randint(100000, 999999)


def generate_loss_curve(epoch, max_epochs, base_loss=2.5, min_loss=0.1):
    """Generate a realistic loss curve that decreases over time with noise"""
    progress = epoch / max_epochs
    base_curve = base_loss * math.exp(-3 * progress) + min_loss

    noise_scale = 0.3 * (1 - progress * 0.7)
    noise = random.gauss(0, noise_scale)

    return max(min_loss * 0.5, base_curve + noise)


def generate_accuracy_curve(epoch, max_epochs, max_acc=0.95, min_acc=0.1):
    """Generate a realistic accuracy curve that increases over time with noise"""
    progress = epoch / max_epochs
    base_curve = max_acc / (1 + math.exp(-6 * (progress - 0.5))) + min_acc

    noise_scale = 0.08 * (1 - progress * 0.5)
    noise = random.gauss(0, noise_scale)

    return max(0, min(max_acc, base_curve + noise))


def generate_grad_norm_curve(epoch, max_epochs):
    """Generate a gradient norm that starts at infinity and decreases to reasonable values"""
    if epoch == 0:
        return float("inf")
    elif epoch == 1:
        return 1000.0
    elif epoch == 2:
        return 100.0
    else:
        progress = (epoch - 2) / (max_epochs - 2)
        base_value = 50 * math.exp(-4 * progress) + 1.0
        noise = random.gauss(0, 0.5)
        return max(0.1, base_value + noise)


for run in range(10):
    wandb.init(
        project=f"fake-training-{PROJECT_ID}",
        name=f"test-run-{10-run}",
        config=dict(
            epochs=EPOCHS,
            learning_rate=0.001,
            batch_size=32,
        ),
    )

    for epoch in range(EPOCHS):
        train_loss = generate_loss_curve(
            epoch,
            EPOCHS,
            base_loss=random.uniform(2.5, 3.5),
            min_loss=random.uniform(0.05, 0.15),
        )
        val_loss = generate_loss_curve(
            epoch,
            EPOCHS,
            base_loss=random.uniform(2.5, 3.5),
            min_loss=random.uniform(0.05, 0.15),
        )

        train_accuracy = generate_accuracy_curve(
            epoch,
            EPOCHS,
            max_acc=random.uniform(0.7, 0.9),
            min_acc=random.uniform(0.1, 0.3),
        )
        val_accuracy = generate_accuracy_curve(
            epoch,
            EPOCHS,
            max_acc=random.uniform(0.7, 0.9),
            min_acc=random.uniform(0.1, 0.3),
        )

        grad_norm = generate_grad_norm_curve(epoch, EPOCHS)

        if epoch > 2 and random.random() < 0.3:
            val_loss *= 1.1
            val_accuracy *= 0.95

        wandb.log(
            {
                "train/loss": round(train_loss, 4),
                "train/accuracy": round(train_accuracy, 4),
                "train/rewards/reward1": random.random(),
                "train/rewards/reward2": random.random(),
                "val/loss": round(val_loss, 4),
                "val/accuracy": round(val_accuracy, 4),
                "grad_norm": grad_norm,
            }
        )

        time.sleep(0.2)

wandb.finish()

[gradio-pr-bot]

🪼 branch checks and previews

•	Name	Status	URL
🦄	Changes	detected!	Details

[gradio-pr-bot]

🦄 change detected

This Pull Request includes changes to the following packages.

Package	Version
trackio	patch

---

• Fix run selection issue

---

✅ Changeset approved by @abidlabs

• Maintainers can remove approval by unchecking this checkbox.

Something isn't right?

• Maintainers can change the version label to modify the version bump.
• If the bot has failed to detect any changes, or if this pull request needs to update multiple packages to different versions or requires a more comprehensive changelog entry, maintainers can update the changelog file directly.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[abidlabs]

Going to merge this in for the release!