wx-ai training create

Syntax

cpdctl wx-ai training create \
[--auto-update-model AUTO-UPDATE-MODEL] \
[--custom CUSTOM] \
[--description DESCRIPTION] \
--name NAME \
[--project-id PROJECT-ID] \
[--prompt-tuning PROMPT-TUNING | --prompt-tuning-base-model PROMPT-TUNING-BASE-MODEL] \
--prompt-tuning-accumulate-steps PROMPT-TUNING-ACCUMULATE-STEPS \
--prompt-tuning-batch-size PROMPT-TUNING-BATCH-SIZE \
--prompt-tuning-init-method PROMPT-TUNING-INIT-METHOD \
[-prompt-tuning-init-text PROMPT-TUNING-INIT-TEXT] \
--prompt-tuning-learning-rate PROMPT-TUNING-LEARNING-RATE \
--prompt-tuning-max-input-tokens PROMPT-TUNING-MAX-INPUT-TOKENS \
--prompt-tuning-max-output-tokens PROMPT-TUNING-MAX-OUTPUT-TOKENS \
--prompt-tuning-num-epochs PROMPT-TUNING-NUM-EPOCHS \
--prompt-tuning-task-id PROMPT-TUNING-TASK-ID \
--prompt-tuning-tuning-type PROMPT-TUNING-TUNING-TYPE \
--prompt-tuning-verbalizer PROMPT-TUNING-VERBALIZER \
[--results-reference RESULTS-REFERENCE | --results-reference-id RESULTS-REFERENCE-ID] \
--results-reference-connection RESULTS-REFERENCE-CONNECTION \
[--results-reference-location RESULTS-REFERENCE-LOCATION] \
--results-reference-type RESULTS-REFERENCE-TYPE \
[--space-id SPACE-ID] \
[--tags TAGS] \
[--training-data-references TRAINING-DATA-REFERENCES]

Extended description

To deploy the tuned model, follow these steps:

1. Provide the details of the base model and parameters for the training in the prompt_tuning object.
2. Create a WML model asset, in a space or a project, by providing a request.json that is similar to the following example.

   curl -X POST "https://{cpd_cluster}/ml/v4/models?version=2024-01-29" \
     -H "Authorization: Bearer <replace with your token>" \
     -H "content-type: application/json" \
     --data '{
        "name": "replace_with_a_meaningful_name",
        "space_id": "replace_with_your_space_id",
        "type": "prompt_tune_1.0",
        "software_spec": {
          "name": "watsonx-textgen-fm-1.0"
        },
        "metrics": [ from the training job ],
        "training": {
          "id": "05859469-b25b-420e-aefe-4a5cb6b595eb",
          "base_model": {
            "model_id": "google/flan-t5-xl"
          },
          "task_id": "generation",
          "verbalizer": "Input: {{input}} Output:"
        },
        "training_data_references": [
          {
            "connection": {
              "id": "20933468-7e8a-4706-bc90-f0a09332b263"
            },
            "id": "file_to_tune1.json",
            "location": {
              "bucket": "wxproject-donotdelete-pr-xeyivy0rx3vrbl",
              "path": "file_to_tune1.json"
            },
            "type": "connection_asset"
          }
        ]
      }'

   Note:

   • If you used the training request field auto_update_model: true, then you can skip this step because the model was saved at the end of the training job.
   • You can use the generated request.json that was stored in the results_reference field, instead of creating the payload for the model. Look for the path in the field entity.results_reference.location.model_request_path.
   • The model type must be prompt_tune_1.0.
   • The software spec name must be watsonx-textgen-fm-1.0.

3. Create a tuned model deployment as described in wx-ai deployment create.

Options

[ {
  "id" : "tune1_data.json",
  "type" : "container",
  "connection" : { },
  "location" : { },
  "schema" : {
    "id" : "t1",
    "name" : "Tasks",
    "fields" : [ {
      "anyKey" : "anyValue"
    }, {
      "anotherAnyKey" : "anotherAnyValue"
    } ],
    "type" : "struct"
  }
} ]

Examples

cpdctl wx-ai training create \
    --name my-prompt-tune-training \
    --results-reference '{"id": "exampleString", "type": "container", "connection": {}, "location": {}}' \
    --space-id exampleString \
    --project-id 12ac4cf1-252f-424b-b52d-5cdd9814987f \
    --description exampleString \
    --tags exampleString,anotherTestString \
    --prompt-tuning '{"base_model": {"model_id": "google/flan-t5-xl"}, "task_id": "classification", "tuning_type": "prompt_tuning", "num_epochs": 30, "learning_rate": 0.4, "accumulate_steps": 3, "verbalizer": "rte { 0 : entailment, 1 : not entailment } {{input}}", "batch_size": 10, "max_input_tokens": 100, "max_output_tokens": 100, "init_method": "text", "init_text": "exampleString"}' \
    --training-data-references '[{"id": "tune1_data.json", "type": "container", "connection": {}, "location": {}, "schema": {"id": "t1", "name": "Tasks", "fields": [{"anyKey": "anyValue"},{"anotherAnyKey": "anotherAnyValue"}], "type": "struct"}}]' \
    --custom '{"anyKey": "anyValue"}' \
    --auto-update-model true

Alternatively, granular options are available for the sub-fields of JSON string options.

cpdctl wx-ai training create \
    --name my-prompt-tune-training \
    --space-id exampleString \
    --project-id 12ac4cf1-252f-424b-b52d-5cdd9814987f \
    --description exampleString \
    --tags exampleString,anotherTestString \
    --training-data-references '[dataConnectionReference]' \
    --custom '{"anyKey": "anyValue"}' \
    --auto-update-model true \
    --results-reference-id exampleString \
    --results-reference-type connection_asset \
    --results-reference-connection dataConnection \
    --results-reference-location '{}' \
    --prompt-tuning-base-model baseModel \
    --prompt-tuning-task-id summarization \
    --prompt-tuning-tuning-type prompt_tuning \
    --prompt-tuning-num-epochs 30 \
    --prompt-tuning-learning-rate 0.4 \
    --prompt-tuning-accumulate-steps 32 \
    --prompt-tuning-verbalizer 'rte { 0 : entailment, 1 : not entailment } {{input}}' \
    --prompt-tuning-batch-size 10 \
    --prompt-tuning-max-input-tokens 100 \
    --prompt-tuning-max-output-tokens 100 \
    --prompt-tuning-init-method text \
    --prompt-tuning-init-text exampleString

Example output

{
  "metadata" : {
    "id" : "6213cf1-252f-424b-b52d-5cdd9814956c",
    "name" : "my-prompt-training",
    "project_id" : "12ac4cf1-252f-424b-b52d-5cdd9814987f",
    "created_at" : "2023-08-04T13:22:47.000Z"
  },
  "entity" : {
    "prompt_tuning" : {
      "base_model" : {
        "model_id" : "google/flan-t5-xl"
      },
      "task_id" : "classification"
    },
    "training_data_references" : [ {
      "id" : "tune1_data.json",
      "location" : {
        "path" : "tune1_data.json"
      },
      "type" : "container"
    } ],
    "auto_update_model" : true,
    "results_reference" : {
      "location" : {
        "path" : "tune1/results",
        "training" : "tune1/results/360c40f7-ac0c-43ca-a95f-1a5421f93b82",
        "training_status" : "tune1/results/360c40f7-ac0c-43ca-a95f-1a5421f93b82/training-status.json",
        "assets_path" : "tune1/results/360c40f7-ac0c-43ca-a95f-1a5421f93b82/assets",
        "model_request_path" : "tune1/results/360c40f7-ac0c-43ca-a95f-1a5421f93b82/assets/c29e7544-dfd0-4427-bc66-20fa6023e2e0/resources/wml_model/request.json"
      },
      "type" : "container"
    },
    "status" : {
      "state" : "completed",
      "running_at" : "2023-08-04T13:22:48.000Z",
      "completed_at" : "2023-08-04T13:22:55.289Z",
      "metrics" : [ {
        "iteration" : 0,
        "ml_metrics" : {
          "loss" : 4.49988
        },
        "timestamp" : "2023-09-22T02:52:03.324Z"
      }, {
        "iteration" : 1,
        "ml_metrics" : {
          "loss" : 3.86884
        },
        "timestamp" : "2023-09-22T02:52:03.689Z"
      }, {
        "iteration" : 2,
        "ml_metrics" : {
          "loss" : 4.05115
        },
        "timestamp" : "2023-09-22T02:52:04.053Z"
      } ]
    }
  }
}