timm Training

Fine-tune any of 1,000+ timm vision models — ViTs, ResNets, EfficientNets, ConvNeXt — on your 3LC tables with per-sample metrics for data debugging and fixing.

1,000+ models

Search and select from the full timm model zoo

Per-sample loss

Track individual losses to find mislabeled or hard examples

Embeddings

Collect UMAP/PaCMAP embeddings for clustering in the Dashboard

Collect mode

Run inference-only to collect metrics without training

1 Select tables

2 Pick model

3 Train

1 Input Tables

Config Name

Train Table URL

Select your training data table. Task type and class names are auto-detected from the table schema.

Validation Table URL (optional — uses train table if empty)

Optional validation set for tracking accuracy during training. If empty, a split of the train table is used.

Use .latest() at train time (resolve to newest table revision)

Detected Task

Columns

Image: -- Label: --

2 timm Settings

Model

Search by name or family (e.g. "resnet", "vit", "efficientnet"). Popular models are shown first.

Mode

Train a model with 3LC metrics collection.

Epochs

Batch Size

Workers

Image Size

Learning Rate

Weight Decay

Optimizer

Scheduler

Warmup Epochs

Auto Augment

Random Erasing

MixUp Alpha

CutMix Alpha

Horizontal Flip

Color Jitter

Pretrained Model URL (optional — fine-tune from existing checkpoint)

Copy best model to run folder (store checkpoint alongside run data)

3 3LC Settings

3LC Project

Run Name

Embeddings Dim

Embeddings Reducer

Keep full (unreduced) embeddings

Collection Start Epoch

Collection Interval

Collect Loss

Sampling Weights

Exclude Zero (Train)

Val Only Collection

Disable Collection