Currently, the only training that can be done through the training container is for the intent classifier. The training code, which is contained inside train_intent_classifier.py runs automatically when running the container with the command:
$ docker compose up training
The training script takes the three datasets that are stored inside training/datasets, then saves the evaluation results from the test and validation sets in a folder eval_results after execution. The fine-tuned model will also be saved in the same training folder.
Important: To train the intent classifier, the container should be run with a GPU available to it.
In case you want to add new samples to the training set to fix intent misclassifications, you should add your samples to the train_set.jsonl file that can be found inside training/datasets. The samples should have the same format as all the other samples inside the jsonl file, for example:
{"system": "Step 3. In a third bowl, use a hand mixer to beat the egg whites
until soft peaks form.",
"user": "Can you go to step 1 please",
"session_id": "",
"intent_pred": "",
"annotation": "step_select(1)"}
The sesssion_id and intent_pred fields can be left empty if the sample is not taken from the system logs.
The container is configured to allow arguments to be passed on the command line if you need to override any of the defaults. Run docker compose run training -h to see the available options. For example, to use a different batch size:
docker compose run training --batch_size 16