Skip to content

Quick Start

SafeTune is a library of safety methods, not a pipeline — you pick one method per task. This page shows the same task (running a train-time defence, or harden) through all three entry points: Python, CLI, and YAML. For the full reference see Usage.

New here? The fastest thing to run is the no-training STEER demo:

python examples/quickstart/quickstart.py

It extracts a refusal direction from contrast prompts, ablates it live on Qwen/Qwen2.5-0.5B-Instruct, and prints the change — the full STEER path, end to end, in about 30 seconds on CPU.

1. Python API

The safetune.runner package wraps each pillar in a uniform Trainer interface. A harden trainer replaces your transformers.Trainer — it is the fine-tuning:

from safetune.runner import harden

# SafeGrad projects the task gradient off the alignment gradient each step,
# so fine-tuning on (possibly contaminated) data does not erode safety.
trainer = harden.SafeGradTrainer(model, tokenizer)
trainer.train(train_dataset, safety_dataset=safety_data)

Every other pillar follows the same shape — for example, inference-time steering with no training:

from safetune.runner import steer

trainer = steer.RefusalDirectionTrainer(model, tokenizer)
wrapped, _ = trainer.calibrate(harmful=harmful_prompts, harmless=harmless_prompts)

2. CLI

After install, the safetune command is available:

safetune list           # print every method, grouped by pillar

# Train with SafeGrad (the default harden method)
safetune train --model Qwen/Qwen2.5-0.5B-Instruct --algo safegrad --epochs 3

# Score a model on a safety benchmark
safetune eval --model Qwen/Qwen2.5-0.5B-Instruct --dataset harmbench

Run safetune --help for the full flag list, or see Usage → CLI.

3. YAML config

Declare all flags in a YAML file and pass it with --config. Any key that is not a standard field is forwarded as a keyword argument to the trainer, so method-specific hyperparameters live in the same file:

# run.yaml
algo: lisa
model: Qwen/Qwen2.5-0.5B-Instruct
epochs: 3
batch_size: 4
train_dataset: beavertails
train_split: 30k_train
output: ./results/lisa

# Method-specific kwargs — forwarded to LisaTrainer(...)
lisa_rho: 0.2
lisa_warmup_steps: 20

safetune train --config run.yaml
safetune train --config run.yaml --epochs 5   # explicit flags override YAML

Next steps

  • Examples — a runnable script and notebook per pillar.
  • Usage — full CLI, YAML, and Python API reference.
  • Getting started — the decision guide for picking a method.