llama-farm Text Classification Skill
Guide for text classification in LlamaFarm using zero-shot and custom-trained approaches.
When to Load
Load this skill when the user:
- Wants to classify or categorize text
- Asks about zero-shot classification
- Needs to train a custom text classifier
- Asks about SetFit or few-shot learning
- Wants to label or tag documents
- Needs sentiment analysis or topic classification
Two Approaches
| Feature | Zero-Shot | Custom (SetFit) |
|---|---|---|
| Training data needed | None | 8-64 examples per class |
| Setup time | Instant | Minutes |
| Accuracy | Good for broad categories | Excellent for specific domains |
| Custom labels | Any labels at inference | Fixed at training time |
| Best for | Exploration, prototyping | Production, domain-specific |
Decision Tree
- No training data available? → Zero-shot
- Labels change frequently? → Zero-shot
- Need high accuracy on specific domain? → Custom SetFit
- Have 8+ examples per class? → Custom SetFit
- Quick prototype? → Start zero-shot, upgrade to SetFit if needed
API Overview
Zero-Shot Classification
| Endpoint | Method | Description |
|---|---|---|
/v1/classify |
POST | Classify text with candidate labels |
Custom Classification (SetFit)
| Endpoint | Method | Description |
|---|---|---|
/v1/classifier/fit |
POST | Train a custom classifier |
/v1/classifier/predict |
POST | Predict with trained classifier |
/v1/classifier/models |
GET | List trained classifiers |
/v1/classifier/models/{name} |
DELETE | Delete a classifier |
Quick Start
Zero-shot classification
curl -X POST http://localhost:14345/v1/classify \
-H "Content-Type: application/json" \
-d '{
"text": "The server crashed and lost all data",
"labels": ["bug", "feature request", "question", "documentation"]
}'
Response:
{
"label": "bug",
"scores": {
"bug": 0.82,
"feature request": 0.08,
"question": 0.06,
"documentation": 0.04
}
}
Train a custom classifier
curl -X POST http://localhost:14345/v1/classifier/fit \
-H "Content-Type: application/json" \
-d '{
"model_name": "ticket_classifier",
"training_data": [
{"text": "App crashes on login", "label": "bug"},
{"text": "Add dark mode", "label": "feature"},
{"text": "How do I reset password?", "label": "question"}
],
"base_model": "sentence-transformers/all-MiniLM-L6-v2",
"num_iterations": 20
}'
Predict with custom classifier
curl -X POST http://localhost:14345/v1/classifier/predict \
-H "Content-Type: application/json" \
-d '{
"model_name": "ticket_classifier",
"text": "The export button does nothing when clicked"
}'
Progressive Disclosure
For detailed guidance:
- zero-shot.md - Pre-trained models, crafting labels, confidence interpretation
- custom-setfit.md - Training workflow, base model selection, evaluation strategies
- model-management.md - Save/load lifecycle, versioning, comparing models