CLI Reference

SafeTune ships a single console script, safetune, that dispatches to each safety pillar. The command is positional — safetune <command> [flags], not a set of subcommand parsers — so global flags apply to every command. The five commands are train (Harden), eval (Evaluate), patch (Recover), unlearn (Unlearn), and list. The CLI covers the common train/recover/unlearn/evaluate paths; steer and interpret, and the programmatic-only harden methods, live in the Python API.

Installation

pip install safetune
safetune list

safetune list prints every registered method grouped by pillar (Harden, Recover, Unlearn) and needs no model — use it to discover the exact --algo alias for a method.

Command Overview

Command	Pillar	Description
train	Harden	Train-time defense: fine-tune a model with a harden trainer and save the hardened weights.
patch	Recover	Weight-space recovery: apply a recover method to a drifted model, optionally using base/aligned references.
unlearn	Unlearn	Forget-set training: build an unlearn trainer for a model, then call it from Python.
eval	Evaluate	Run safety/utility benchmarks against a model and print per-benchmark metrics.
list	—	Print every available method per pillar. No model required.

safetune train

Harden a model by fine-tuning it with a train-time defense. The trainer named by --algo is looked up in HARDEN_REGISTRY; the model is loaded, the training dataset (--train-dataset / --train-split, default BeaverTails 30k_train) is tokenized, and the hardened weights are written to --output.

safetune train \
  --model Qwen/Qwen2.5-0.5B-Instruct \
  --algo lisa \
  --train-dataset beavertails \
  --epochs 3 --batch-size 4 --lr 2e-5 \
  --output ./results/lisa

Relevant flags: --algo, --model, --train-dataset, --train-split, --epochs, --batch-size, --lr, --precision, --output, --wandb.

Programmatic-only harden methods

A few harden methods — cst, mart, deeprefusal, antibody — need method-specific data/config that the uniform CLI train contract can't supply. Running e.g. safetune train --algo cst prints a hint pointing at the Python API (from safetune.runner.harden import CSTTrainer) and exits. Use these from Python; see the harden guide.

safetune eval

Evaluate a model against one or more benchmarks and print each benchmark's metrics. Pass --dataset as a comma-separated list of benchmark names; omit it to run the default suite. --eval-backend selects the generation backend (vllm or hf); the default is auto — vLLM when installed, otherwise transformers.

safetune eval \
  --model Qwen/Qwen2.5-0.5B-Instruct \
  --dataset harmbench,advbench \
  --eval-backend vllm

Relevant flags: --model, --dataset, --eval-backend.

safetune patch

Recover safety in a drifted model via weight-space editing. The trainer named by --algo is looked up in RECOVER_REGISTRY. Many recover methods need reference checkpoints: --base supplies the base model and --aligned the aligned model used to extract the safety delta. The command applies the patch in memory and reminds you to call trainer.save_checkpoint() from Python to persist it.

safetune patch \
  --model ./drifted \
  --algo resta \
  --base ./base

Relevant flags: --algo, --model, --base, --aligned.

safetune unlearn

Build an unlearn (forget-set training) trainer for a model. The trainer named by --algo is looked up in UNLEARN_REGISTRY. The CLI constructs the trainer and prints readiness; the actual forget/retain step runs in Python via trainer.unlearn(forget=..., retain=...), since it needs the forget and retain sets.

safetune unlearn \
  --model ./model \
  --algo rmu \
  --epochs 1 --lr 5e-5

Relevant flags: --algo, --model, --epochs, --batch-size, --lr, --precision.

safetune list

Print every registered method for each pillar (Harden, Recover, Unlearn) as alias → TrainerClass, followed by example invocations. Requires no --model.

safetune list

Global Flags

All flags are global — they parse regardless of command, though not every command reads every flag. --model is required for every command except list.

Flag	Default	Description
--config	—	Path to a YAML config file. Values are overridden by explicit CLI flags.
--algo	safegrad	Method (alias) within the pillar; see safetune list.
--model	—	Model path or HF hub ID. Required except for list.
--dataset	—	Dataset or comma-separated benchmarks for eval.
--output, --output-dir	./results	Output directory.
--epochs	1	Training epochs.
--batch-size	1	Batch size.
--lr, --learning-rate	5e-5	Learning rate / alpha.
--base	—	Base model path (for patch/recover).
--aligned	—	Aligned model path (for patch/recover).
--precision	bf16	Compute precision: fp16, bf16, or fp32.
--train-dataset	beavertails	Training dataset: beavertails or any HF dataset id.
--train-split	—	Split for --train-dataset (default: 30k_train for beavertails, else train).
--eval-backend	auto	Generation backend for eval: vllm or hf (auto: vLLM if installed, else hf).
--wandb	off	Enable Weights & Biases logging.

YAML Configuration

Instead of passing every flag, describe a run declaratively with --config. The YAML maps onto SafeTuneConfig: recognized keys (command, algo, model, base, aligned, output, epochs, batch_size, lr, precision, train_dataset, train_split, dataset, eval_backend, ...) become standard fields, and any unrecognized key is forwarded to the trainer as a method-specific keyword argument (e.g. lisa_rho, rank, inner_steps).

# lisa.yaml
command: train
algo: lisa
model: Qwen/Qwen2.5-0.5B-Instruct
epochs: 3
batch_size: 4
lr: 2e-5
train_dataset: beavertails
train_split: 30k_train
output: ./results/lisa
# method-specific kwargs — forwarded to the trainer
lisa_rho: 0.2
lisa_warmup_steps: 20

safetune --config lisa.yaml train

Precedence: --config values seed the parser defaults, and explicit CLI flags override them. So the config supplies the baseline and any flag on the command line wins:

# Uses lisa.yaml, but overrides the model on the command line
safetune --config lisa.yaml train --model other/model