Skip to content

CLI Reference

The tobira command provides subcommands for server management, diagnostics, evaluation, and monitoring.

Installation

pip install tobira[serving]

Getting Started

tobira init

Auto-detect your MTA and generate integration configuration files.

tobira init [--mta MTA] [--output-dir DIR]

The setup wizard:

  1. Detects running MTAs (rspamd, SpamAssassin, Haraka, Postfix)
  2. Generates MTA-specific plugin configuration
  3. Provides step-by-step installation instructions

tobira serve

Start the FastAPI API server.

tobira serve [--config CONFIG] [--host HOST] [--port PORT] [--backend BACKEND]
Option Default Description
--config — Path to TOML configuration file
--host 127.0.0.1 Server bind address
--port 8000 Server bind port
--backend — Backend type (overrides config)
--model-path — Model file path (overrides config)

tobira doctor

Run diagnostic checks on your tobira installation.

tobira doctor [--config CONFIG]

Checks performed:

  • Configuration file validity
  • Backend model loading
  • API server connectivity
  • MTA plugin activation status

tobira demo

Launch a full demonstration environment using Docker Compose.

tobira demo [--backend BACKEND]

Starts the API server, MTA plugins, and monitoring stack for interactive evaluation. Useful for trying tobira without configuring your production MTA.

Training & Evaluation

tobira train

Fine-tune a BERT/DeBERTa model on labeled email data.

tobira train --config CONFIG --data DATA --output OUTPUT
Option Description
--config Configuration file path
--data Labeled data file (CSV or JSONL)
--output Output directory for trained model

Supports automatic train/val/test split, ONNX export, and training metrics persistence.

tobira evaluate

Evaluate a backend against labeled test data.

tobira evaluate --data DATA [--backend BACKEND] [--config CONFIG]
Option Description
--data Path to labeled data file (JSONL)
--backend Backend type to evaluate
--config Configuration file path

Outputs: accuracy, precision, recall, F1 score, confusion matrix, and PR curve.

tobira monitor

Analyze prediction metrics and suggest improvements.

tobira monitor [--config CONFIG] [--watch] [--interval SECONDS]
Option Default Description
--config — Configuration file path
--watch — Enable daemon mode (continuous monitoring)
--interval 3600 Check interval in seconds (daemon mode)

Analysis includes:

  • Score distribution shift detection (PSI / KS test)
  • FP/FN rate trends
  • Threshold tuning suggestions
  • Backend upgrade recommendations

Advanced

tobira distill

Run knowledge distillation from a teacher model to a student model.

tobira distill --teacher TEACHER --student STUDENT --data DATA
Option Description
--teacher Teacher model path or name
--student Student model path or name
--data Training data path

tobira ab-test

Run A/B testing between model variants in production.

tobira ab-test --config CONFIG

Supports random and hash-based traffic splitting with variant analytics.

tobira active-learning

Manage the active learning queue for uncertainty-based sample selection.

tobira active-learning --config CONFIG [--strategy STRATEGY]
Option Default Description
--config — Configuration file path
--strategy entropy Sampling strategy (entropy, bald, margin)

tobira hub-push / tobira hub-pull

Manage models on HuggingFace Hub.

# Push a model to HuggingFace Hub
tobira hub-push --model-path MODEL --repo REPO

# Pull a model from HuggingFace Hub
tobira hub-pull --repo REPO --output-dir DIR

tobira milter

Start the Postfix milter daemon.

tobira milter --config CONFIG

See Postfix milter integration for configuration details.