Model-training CLI

safe-train trains or fine-tunes a SAFE generative model using the supported Transformers 5 stack.

The tokenizer and dataset are required for training. The output directory is managed by the Transformers trainer:

safe-train \
  --tokenizer path/to/tokenizer.json \
  --dataset path/to/dataset \
  --output_dir path/to/output \
  --do_train \
  --max_steps 1000

To start from an existing checkpoint, add --model_path. To use a custom model configuration, add --config; otherwise SAFE uses its packaged GPT-2 configuration.

Common options include:

Option	Purpose
--tokenizer	Required tokenizer JSON file, local tokenizer directory, or compatible Hub identifier.
--dataset	Required local Datasets directory or Hugging Face dataset identifier.
--model_path	Optional checkpoint used to initialize the model.
--config	Optional Transformers configuration; defaults to SAFE's packaged configuration.
--streaming	Stream the dataset instead of loading it eagerly.
--text_column	Dataset column containing SAFE strings; defaults to inputs.
--include_descriptors	Enable the auxiliary property head.
--property_column	Dataset column containing descriptor targets.
--model_max_length	Token sequence limit; defaults to 1024.
--wandb_project	Weights & Biases project name. Set an empty value to disable W&B reporting.
--do_train, --do_eval, --do_predict	Select the trainer operations to execute.

All standard transformers.TrainingArguments are also available. Run the installed command to see the exact options for the pinned Transformers release:

safe-train --help

Model-only linker generation uses a reviewed, commit-pinned constrained-beam backend because that algorithm moved out of the main Transformers package. See Migrating to SAFE 0.2.0 for details and offline configuration.