Using CLI

We provide a ruprompts-train endpoint, that implements the training loop described in Walkthrough, parameterized by the config described in Config Structure.

Info

Note that the endpoint is not installed by default. To install it, you should add an extra:

pip install ruprompts[hydra]

See Installation for details.

The endpoint behaves as a standard Hydra application, and its parameters should be passed in no hyphen format.

Examples

Training a task defined in Config Structure:

ruprompts-train \
    task=detoxification training.run_name=detox-tensor-linear-lr-1e-1 \
    prompt_provider=tensor training.learning_rate=1e-1 scheduler=linear_schedule_with_warmup

Defining task purely in command line:

ruprompts-train \
    task=text2text training.run_name=very-custom-run \
    task.task_name=detoxification \
    prompt_provider=tensor \
    +dataset=from_jsonl \
    +dataset.data_files.train=/path/to/train.jsonl \
    +dataset.data_files.validation=/path/to/validation.jsonl \
    prompt_format.template='"<P*60>{toxic}<P*20>"' \
    preprocessing.target_field=polite \
    preprocessing.truncation_field=toxic \
    preprocessing.max_tokens=1792

or with dataset from HF Hub:

ruprompts-train \
    task=text2text training.run_name=summarization-mlsum \
    task.task_name=summarization \
    prompt_provider=tensor \
    +dataset=default \
    +dataset.path=mlsum \
    +dataset.name=ru \
    prompt_format.template='"<P*60>{text}<P*20>"' \
    preprocessing.target_field=summary \
    preprocessing.truncation_field=text \
    preprocessing.max_tokens=1792

Working directory behaviour

By default ruprompts-train creates a new working directory for each run, grouping them by task name and separating debug runs into the debug folder. Thus, for non-debug runs the workdir is switched to ./outputs/{task_name}/{datetime} relatively to the directory where the endpoint was called, and to ./outputs/debug/{task_name}/{datetime} for debug runs respectively.

The working directory contains all the logs and checkpoints. For example, loading a prompt after running ruprompts-train will be done with something like Prompt.from_pretrained("./outputs/debug/detoxification/20211231_235959").

However, switching workdir in runtime makes using relative paths in configs (e.g. for data files) tricky and unreliable. For safety use only absolute paths.