# Copyright 2023 The Kubeflow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Vertex Gen AI Evaluation for Text Generation/QA/Summarization tasks."""
from typing import Dict, List, NamedTuple
from google_cloud_pipeline_components._implementation.model_evaluation import LLMEvaluationPreprocessorOp
from google_cloud_pipeline_components._implementation.model_evaluation import LLMEvaluationTextGenerationOp
from google_cloud_pipeline_components._implementation.model_evaluation import ModelImportEvaluationOp
from google_cloud_pipeline_components._implementation.model_evaluation import ModelNamePreprocessorOp
from google_cloud_pipeline_components.types.artifact_types import VertexModel
from google_cloud_pipeline_components.v1.batch_predict_job import ModelBatchPredictOp
from kfp import dsl
# pylint: disable=unused-argument, unexpected-keyword-arg
_PIPELINE_NAME = 'evaluation-llm-text-generation-pipeline'
[docs]@dsl.pipeline(name=_PIPELINE_NAME)
def evaluation_llm_text_generation_pipeline( # pylint: disable=dangerous-default-value
project: str,
location: str,
batch_predict_gcs_source_uris: List[str],
batch_predict_gcs_destination_output_uri: str,
service_account: str,
model_name: str = 'publishers/google/models/text-bison@002',
evaluation_task: str = 'text-generation',
input_field_name: str = 'input_text',
target_field_name: str = 'output_text',
batch_predict_instances_format: str = 'jsonl',
batch_predict_predictions_format: str = 'jsonl',
batch_predict_model_parameters: Dict[str, str] = {},
enable_row_based_metrics: bool = False,
machine_type: str = 'e2-standard-4',
network: str = '',
encryption_spec_key_name: str = '',
evaluation_display_name: str = 'evaluation-llm-text-generation-pipeline-{{$.pipeline_job_uuid}}',
) -> NamedTuple(
'outputs', evaluation_metrics=dsl.Metrics, evaluation_resource_name=str
):
# fmt: off
"""LLM Text Generation Evaluation pipeline.
This pipeline supports evaluating large language models, publisher or managed
models, performing the following generative tasks: `summarization`, `question-answering`, and `text-generation`.
Args:
project: Required. The GCP project that runs the pipeline components.
location: Required. The GCP region that runs the pipeline components.
batch_predict_gcs_source_uris: Required. Google Cloud Storage URI(s) to your eval dataset instances data to run batch prediction on. The instances data should also contain the ground truth (target) data, used for evaluation. May contain wildcards. For more information on [wildcards](https://cloud.google.com/storage/docs/gsutil/addlhelp/WildcardNames). For more details about this [input config](https://cloud.google.com/vertex-ai/docs/reference/rest/v1/projects.locations.batchPredictionJobs#InputConfig). The content of gcs source files should be preset to one of the following formats:
1) Prediction & Evaluation Dataset format, guaranteeing "prompt" and "ground_truth" attributes are included
{
"prompt": "your input/prompt text",
"ground_truth": "your ground truth output text"
}
or
2) Tuning Dataset format, guaranteeing "input_text" and "output_text" attributes are included.
{
"input_text": "your input/prompt text",
"output_text": "your ground truth output text"
}
batch_predict_gcs_destination_output_uri: Required. The Google Cloud Storage location of the directory where the eval pipeline output is to be written to.
service_account: Required. Sets the default service account for workload run-as account. The service account running the pipeline (https://cloud.google.com/vertex-ai/docs/pipelines/configure-project#service-account) submitting jobs must have act-as permission on this run-as account.
model_name: The Model name used to run evaluation. Must be a publisher Model or a managed Model sharing the same ancestor location. Starting this job has no impact on any existing deployments of the Model and their resources.
evaluation_task: The task that the large language model will be evaluated on. The evaluation component computes a set of metrics relevant to that specific task. Currently supported tasks are: `summarization`, `question-answering`, `text-generation`.
input_field_name: The field name of the input eval dataset instances that contains the input prompts to the LLM.
target_field_name: The field name of the eval dataset instance that contains an example reference text response. Alternatively referred to as the ground truth (or ground_truth_column) field. If not set, defaulted to `output_text`.
batch_predict_instances_format: The format in which instances are given, must be one of the Model's supportedInputStorageFormats. Only "jsonl" is currently supported. For more details about this input config, see https://cloud.google.com/vertex-ai/docs/reference/rest/v1/projects.locations.batchPredictionJobs#InputConfig.
batch_predict_predictions_format: The format in which Vertex AI gives the predictions. Must be one of the Model's supportedOutputStorageFormats. Only "jsonl" is currently supported. For more details about this output config, see https://cloud.google.com/vertex-ai/docs/reference/rest/v1/projects.locations.batchPredictionJobs#OutputConfig.
batch_predict_model_parameters: A map of parameters that govern the predictions. Some acceptable parameters include: maxOutputTokens, topK, topP, and temperature.
enable_row_based_metrics: Flag of if row based metrics is enabled, default value is false.
machine_type: The machine type of this custom job. If not set, defaulted to `e2-standard-4`. More details: https://cloud.google.com/compute/docs/machine-resource
network: The full name of the Compute Engine network to which the job should be peered. For example, `projects/12345/global/networks/myVPC`. Format is of the form `projects/{project}/global/networks/{network}`. Where `{project}` is a project number, as in `12345`, and `{network}` is a network name, as in `myVPC`. To specify this field, you must have already configured VPC Network Peering for Vertex AI (https://cloud.google.com/vertex-ai/docs/general/vpc-peering). If left unspecified, the job is not peered with any network.
encryption_spec_key_name: Customer-managed encryption key options. If set, resources created by this pipeline will be encrypted with the provided encryption key. Has the form: `projects/my-project/locations/my-location/keyRings/my-kr/cryptoKeys/my-key`. The key needs to be in the same region as where the compute resource is created.
evaluation_display_name: The display name of the uploaded evaluation resource to the Vertex AI model.
Returns:
evaluation_metrics: Metrics Artifact for LLM Text Generation.
evaluation_resource_name: If run on a user's managed VertexModel, the imported evaluation resource name. Empty if run on a publisher model.
"""
# fmt: on
outputs = NamedTuple(
'outputs',
evaluation_metrics=dsl.Metrics,
evaluation_resource_name=str,
)
preprocessed_model_name = ModelNamePreprocessorOp(
project=project,
location=location,
model_name=model_name,
service_account=service_account,
)
get_vertex_model_task = dsl.importer(
artifact_uri=(
f'https://{location}-aiplatform.googleapis.com/v1/{preprocessed_model_name.outputs["processed_model_name"]}'
),
artifact_class=VertexModel,
metadata={
'resourceName': preprocessed_model_name.outputs[
'processed_model_name'
]
},
)
get_vertex_model_task.set_display_name('get-vertex-model')
eval_dataset_preprocessor_task = LLMEvaluationPreprocessorOp(
project=project,
location=location,
gcs_source_uris=batch_predict_gcs_source_uris,
input_field_name=input_field_name,
machine_type=machine_type,
service_account=service_account,
network=network,
encryption_spec_key_name=encryption_spec_key_name,
)
batch_predict_task = ModelBatchPredictOp(
project=project,
location=location,
model=get_vertex_model_task.outputs['artifact'],
job_display_name='evaluation-batch-predict-{{$.pipeline_job_uuid}}-{{$.pipeline_task_uuid}}',
gcs_source_uris=eval_dataset_preprocessor_task.outputs[
'preprocessed_gcs_source_uris'
],
instances_format=batch_predict_instances_format,
predictions_format=batch_predict_predictions_format,
gcs_destination_output_uri_prefix=batch_predict_gcs_destination_output_uri,
model_parameters=batch_predict_model_parameters,
encryption_spec_key_name=encryption_spec_key_name,
)
eval_task = LLMEvaluationTextGenerationOp(
project=project,
location=location,
evaluation_task=evaluation_task,
target_field_name=f'instance.{target_field_name}',
predictions_format=batch_predict_predictions_format,
enable_row_based_metrics=enable_row_based_metrics,
joined_predictions_gcs_source=batch_predict_task.outputs[
'gcs_output_directory'
],
machine_type=machine_type,
service_account=service_account,
network=network,
encryption_spec_key_name=encryption_spec_key_name,
)
with dsl.If(enable_row_based_metrics == True):
import_evaluation_task_with_row_based_metrics = ModelImportEvaluationOp(
metrics=eval_task.outputs['evaluation_metrics'],
row_based_metrics=eval_task.outputs['row_based_metrics'],
model=get_vertex_model_task.outputs['artifact'],
problem_type=evaluation_task,
dataset_type=batch_predict_predictions_format,
dataset_paths=batch_predict_gcs_source_uris,
display_name=evaluation_display_name,
)
with dsl.Else():
import_evaluation_task = ModelImportEvaluationOp(
metrics=eval_task.outputs['evaluation_metrics'],
model=get_vertex_model_task.outputs['artifact'],
problem_type=evaluation_task,
dataset_type=batch_predict_predictions_format,
dataset_paths=batch_predict_gcs_source_uris,
display_name=evaluation_display_name,
)
oneof = dsl.OneOf(
import_evaluation_task_with_row_based_metrics.outputs[
'evaluation_resource_name'
],
import_evaluation_task.outputs['evaluation_resource_name'],
)
return outputs(
evaluation_metrics=eval_task.outputs['evaluation_metrics'],
evaluation_resource_name=oneof,
)