Model Finetuning

Finetune pre-trained models on your own data to achieve high accuracy on your specific task. Start from a foundation model and adapt it to your domain.

What is Finetuning?

Finetuning takes a model that was pre-trained on a large dataset and continues training it on your smaller, task-specific dataset. This gives you:

Better accuracy than training from scratch
Less data required (hundreds instead of millions of examples)
Faster training (minutes to hours instead of days)
Domain expertise the base model didn’t have

graph LR
    A[Pre-trained Model
General Knowledge] --> B[Your Dataset
Domain-Specific]
    B --> C[Finetuned Model
Your Task Expert]

    style A fill:#e0f2fe
    style B fill:#fef3c7
    style C fill:#d1fae5

When to Finetune

Scenario	Recommendation
Standard task, enough labeled data	Finetune a pre-trained model
Very specific domain (medical, legal)	Finetune with domain data
Limited labeled data (<100 examples)	Use few-shot or consider LLM
Need maximum accuracy	Finetune + hyperparameter tuning
Need fast inference	Finetune small model or distill

Guide Sections

1. Prepare Data 2. Choose Base Model 3. Configure Training 4. Train & Monitor 5. Evaluate Results 6. Iterate & Improve

Quick Start

1. Prepare Your Dataset

from seeme import Client

client = Client()

## Create a dataset
dataset = client.create_dataset(
    name="Product Defect Classifier",
    task_type="image_classification"
)

version = client.create_dataset_version(
    dataset_id=dataset.id,
    name="v1"
)

# Create splits
train_split = client.create_split(version_id=version.id, name="train")
val_split = client.create_split(version_id=version.id, name="validation")

# Create labels
labels = ["good", "scratch", "dent", "discoloration"]
for label_name in labels:
    client.create_label(version_id=version.id, name=label_name)

# Upload and label data
for image_path, label in training_data:
    item = client.create_dataset_item(
        version_id=version.id,
        split_id=train_split.id,
        file_path=image_path
    )
    client.annotate(
        dataset_id=dataset.id,
        dataset_version_id=version.id,
        annotation=Annotation(
            item_id=item.id,
            split_id=train_split.id,
            label_id=label
        )
    )

2. Start Training

# Create finetuning job
job = client.create_job(
    dataset_id=dataset.id,
    version_id=version.id,
    name="Defect Classifier v1",
    job_type="finetune",
    config={
        "base_model": "efficientnet_b0",  # Pre-trained model
        "epochs": 20,
        "learning_rate": 0.001,
        "batch_size": 32,
        "image_size": 224,
        "freeze_layers": "early",  # Freeze early layers
        "augmentation": {
            "horizontal_flip": True,
            "rotation": 15,
            "brightness": 0.2
        }
    }
)

print(f"Training job started: {job.id}")

3. Monitor Progress

import time

while job.status in ["pending", "running"]:
    time.sleep(30)
    job = client.get_job(job.id)

    if job.metrics:
        print(f"Epoch {job.metrics.get('epoch', '?')}: "
              f"loss={job.metrics.get('loss', '?'):.4f}, "
              f"accuracy={job.metrics.get('accuracy', '?'):.2%}")

print(f"Training complete: {job.status}")
if job.status == "completed":
    model = client.get_model(job.model_id)
    print(f"Model accuracy: {model.metrics['accuracy']:.2%}")

Supported Task Types

Task	Base Models	Data Required
Image Classification	EfficientNet, ResNet, MobileNet, ViT	100-1000+ images per class
Object Detection	YOLOv4, YOLOv5, Faster R-CNN	100-500+ annotated images
Text Classification	BERT, DistilBERT, RoBERTa	100-1000+ examples per class
Named Entity Recognition	spaCy, BERT-NER	500-2000+ annotated sentences
Tabular	XGBoost, LightGBM, CatBoost	1000+ rows

Finetuning vs. Other Approaches

Approach	When to Use	Data Needed	Training Time
Finetuning	Task-specific accuracy needed	100-10,000	Minutes-Hours
Training from scratch	Unique architecture needed	10,000+	Hours-Days
Transfer learning	Similar task to base model	50-500	Minutes
Few-shot (LLM)	Very limited data	5-50	None
Distillation	Need smaller/faster model	Labeled by teacher	Hours

Best Practices

Start with a good base model - Choose one trained on similar data
Use the right learning rate - Lower than training from scratch (0.0001-0.001)
Freeze early layers - Keep general features, train task-specific ones
Augment your data - Especially if dataset is small
Monitor for overfitting - Watch validation loss diverge from training loss
Save checkpoints - Keep models from different epochs
Test on held-out data - Never evaluate on training data

Next Steps

Start with Prepare Data to get your dataset ready for finetuning.