Model Registry

Current Models

Model	Purpose	Languages	Accuracy
clapAI/modernBERT-large-multilingual-sentiment	Sentiment classification	16+	98.5% (IMDB)
AnasAlokla/multilingual_go_emotions_V1.1	Emotion detection (28 classes)	6+	88.6% (GoEmotions)
papluca/xlm-roberta-base-language-detection	Language identification	20+	N/A
unitary/toxic-bert	Toxicity screening	English	N/A

Emotion Classes (28)

Positive: admiration, amusement, approval, caring, desire, excitement, gratitude, joy, love, optimism, pride, relief

Negative: anger, annoyance, disappointment, disapproval, disgust, embarrassment, fear, grief, nervousness, remorse, sadness

Neutral: neutral, realization, confusion, curiosity, surprise

Model Versioning

Models are versioned in MLflow. The "champion" alias points to the best performing model.

# View models
open http://localhost:5000

# Promote a model
docker compose exec producer python scripts/promote_model.py

# Hot-swap in production
curl -X POST http://localhost:8000/reload
curl -X POST http://localhost:8000/reload -d '{"version": 5}'

A/B Testing

# Start A/B test: 20% traffic to version 3
curl -X POST http://localhost:8000/ab-test \
  -H "Content-Type: application/json" \
  -d '{"challenger_version": 3, "split": 0.2}'

# Check status
curl http://localhost:8000/ab-test

# Stop A/B test
curl -X DELETE http://localhost:8000/ab-test

Training a New Model

# CI mode (fast, 2K samples)
make train

# Full training (50K samples)
docker compose exec -e TRAINING_MODE=FULL producer ...

# Try a different model
docker compose exec -e HF_MODEL_NAME=textattack/bert-base-uncased-imdb producer ...

Tested Models

Version	Model	Accuracy	Notes
V1	distilbert-sst-2	89.6%	English only, fast
V3	lvwerra/distilbert-imdb	95.0%	English, IMDB-tuned
V5	textattack/bert-imdb	95.5%	English, BERT-based
V9	nlptown/bert-multilingual	90.5%	6 languages, star ratings
V10	modernBERT-large-multilingual	98.5%	16+ languages, best overall