ctrl

Control Tokens

To train a CTRL model, you need to provide control tokens for each utterance in training and validation data.

Requirements

Computing control tokens requires additional dependencies:

pip install -r models/ctrl/ctrl_requirements.txt

Run

To determine control tokens for a given dataset, run the following command:

python models/ctrl/ctrl_data.py --split train --output_dir /path/to/output

Set CUDA_VISIBLE_DEVICES if you wish to run the command to GPU(s).

Available arguments for the above command are:

• --input_file (optional): Path to the input file. If not provided, data will be loaded from Huggingface's datasets.
• --output_dir: Path to the output directory. Optional if --input_file refers to a file.
• --split: Split of the data to use. One of train or valid.
• --nli_model: NLI model for predicting entailment labels.
• --per_device_batch_size: Batch size per device.
• --max_length: Maximum sequence length.