ML Training

The Designer includes built-in tools for training classifiers and anomaly detection models — no ML expertise required. These run on the Universal Runtime.

Text Classifiers

Train a text classifier using SetFit (few-shot fine-tuning). Great for sentiment analysis, ticket routing, content categorization, and more.

How It Works

1. Choose a base model — select an embedder model (default: all-MiniLM-L6-v2)
2. Add training data — enter text/label pairs, or use a sample dataset
3. Train — click Train and watch progress in real-time
4. Test — enter text and see predictions with confidence scores

Available Base Models

Model	Dimensions	Notes
all-MiniLM-L6-v2	384	Default, fast, good general purpose
bge-small-en-v1.5	384	Strong English performance
bge-base-en-v1.5	768	Larger, more accurate
bge-large-en-v1.5	1024	Best accuracy, slower
bge-m3	1024	Multilingual support
e5-base-v2	768	Good for retrieval tasks
e5-large-v2	1024	Larger variant

Sample Datasets

Built-in sample datasets to get started quickly:

• Sentiment analysis — 3 classes, 200 examples (positive/negative/neutral)
• Additional domain-specific samples available

Managing Trained Models

The Trained Models view lists all your classifier models with:

• Model name and version
• Training timestamp
• Number of classes and examples
• Actions: load, test, delete

Anomaly Detection

Train anomaly detection models using PyOD backends. Useful for fraud detection, system monitoring, quality control, and any scenario where you need to identify outliers.

Backends

12 PyOD backends organized by category:

Category	Backends	Best for
Fast (Recommended)	ECOD, HBOS, COPOD	General purpose, parameter-free
Legacy (Well-Tested)	Isolation Forest, LOF, KNN, OCSVM	Traditional ML approaches
Deep Learning	AutoEncoder, VAE	Complex patterns
Ensemble	SUOD, LSCP	Combining multiple detectors

Training Flow

1. Select a backend — ECOD recommended for most cases
2. Configure features — define feature columns with encoding types and normalization
3. Add training data — paste text/CSV or use table input mode
4. Set threshold — contamination ratio (default varies by backend)
5. Train — model trains and shows results

Feature Configuration

Each feature column supports:

• Encoding types — numeric, one-hot, label, ordinal, binary, frequency
• Normalization — standard, min-max, robust, none

Streaming Anomaly Detection

For real-time monitoring, the streaming mode provides:

• Mode toggle — switch between batch and streaming detection
• Status panel — connection status, events processed, anomalies detected
• Results chart — live visualization of anomaly scores over time
• Mode panel — configure streaming parameters

API Routes

Action	Method	Route
Train classifier	POST	/v1/ml/classifier/fit
Predict (classifier)	POST	/v1/ml/classifier/predict
List classifier models	GET	/v1/ml/classifier/models
Train anomaly model	POST	/v1/ml/anomaly/fit
Score anomaly	POST	/v1/ml/anomaly/score
List anomaly models	GET	/v1/ml/anomaly/models

Route

/chat/models/train/classifier/new
/chat/models/train/classifier/:id
/chat/models/train/anomaly/new
/chat/models/train/anomaly/:id