# Copyright 2024 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.
"""Defines an RLAIF Kubeflow pipeline.
This pipeline build preference data from AI feedback, trains a
reward model and performs reinforcement learning.
"""
from typing import NamedTuple, Optional
from google_cloud_pipeline_components import _placeholders
from google_cloud_pipeline_components._implementation.llm import online_evaluation_pairwise
from google_cloud_pipeline_components._implementation.llm import preference_data_formatter
from google_cloud_pipeline_components.preview.llm.infer import component as infer
from google_cloud_pipeline_components.preview.llm.rlhf import component as rlhf
import kfp
class PipelineOutput(NamedTuple):
model_resource_name: str
endpoint_resource_name: str
[docs]@kfp.dsl.pipeline(
name='rlaif_pipeline',
description='Performs reinforcement learning from AI feedback.',
)
def rlaif_pipeline(
prompt_dataset: str,
preference_prompt_dataset: str,
large_model_reference: str,
model_display_name: Optional[str] = None,
prompt_sequence_length: int = 512,
target_sequence_length: int = 64,
large_model_a_reference: str = 'text-bison@001',
large_model_b_reference: str = 't5-small',
reward_model_learning_rate_multiplier: float = 1.0,
reinforcement_learning_rate_multiplier: float = 1.0,
reward_model_train_steps: int = 1000,
reinforcement_learning_train_steps: int = 1000,
kl_coeff: float = 0.1,
sampling_strategy: str = 'temperature_sampling',
instruction: Optional[str] = None,
eval_dataset: Optional[str] = None,
project: str = _placeholders.PROJECT_ID_PLACEHOLDER,
accelerator_type: str = 'GPU',
location: str = _placeholders.LOCATION_PLACEHOLDER,
tensorboard_resource_id: Optional[str] = None,
) -> PipelineOutput:
# fmt: off
"""Performs reinforcement learning from AI feedback.
At the moment, it only supports summarization task type.
Args:
prompt_dataset: Cloud storage path to an unlabled JSONL dataset that contains prompts. Text datasets must contain an `input_text` field that contains the prompt. Chat datasets must contain at least 1 message in a `messages` field. Each message must be valid JSON that contains `author` and `content` fields, where valid `author` values are `user` and `assistant` and `content` must be non-empty. Each row may contain multiple messages, but the first and last author must be the `user`. An optional `context` field may be provided for each example in a chat dataset. If provided, the `context` will preprended to the message `content`. The `instruction` serves as the default context. (Useful if most messages use the same system-level context.) Any context provided in the example will override the default value.
preference_prompt_dataset: The prompt dataset used for two models' inferences to build the side by side comparison AI feedback. large_model_reference: Name of the base model. Supported values are `text-bison@001`, `t5-small`, `t5-large`, `t5-xl` and `t5-xxl`. `text-bison@001` and `t5-small` are supported in `us-central1` and `europe-west4`. `t5-large`, `t5-xl` and `t5-xxl` are only supported in `europe-west4`.
model_display_name: Name of the fine-tuned model shown in the Model Registry. If not provided, a default name will be created.
prompt_sequence_length: Maximum tokenized sequence length for input text. Higher values increase memory overhead. This value should be at most 8192. Default value is 512.
target_sequence_length: Maximum tokenized sequence length for target text. Higher values increase memory overhead. This value should be at most 1024. Default value is 64.
large_model_a_reference: Name of a predefined model A for side by side comparison to build the AI feedback dataset. By default, it uses `text-bison@001`. The valid values are `t5-small`, `t5-large`, `t5-xl`, `t5-xxl`, `text-bison@001`, `llama-2-7b`, `llama-2-13b`.
large_model_b_reference: Name of a predefined model B for side by side comparison to build the AI feedback dataset. By default, it uses `t5-small`. The valid values are `t5-small`, `t5-large`, `t5-xl`, `t5-xxl`, `text-bison@001`, `llama-2-7b`, `llama-2-13b`.
reward_model_learning_rate_multiplier: Constant used to adjust the base learning rate used when training a reward model. Multiply by a number > 1 to increase the magnitude of updates applied at each training step or multiply by a number < 1 to decrease the magnitude of updates. Default value is 1.0.
reinforcement_learning_rate_multiplier: Constant used to adjust the base learning rate used during reinforcement learning. Multiply by a number > 1 to increase the magnitude of updates applied at each training step or multiply by a number < 1 to decrease the magnitude of updates. Default value is 1.0.
reward_model_train_steps: Number of steps to use when training a reward model. Default value is 1000.
reinforcement_learning_train_steps: Number of reinforcement learning steps to perform when tuning a base model. Default value is 1000.
kl_coeff: Coefficient for KL penalty. This regularizes the policy model and penalizes if it diverges from its initial distribution. If set to 0, the reference language model is not loaded into memory. Default value is 0.1.
sampling_strategy: The strategy used to candidates for AI feedback. Default is temperature_sampling. Valid values are greedy, temperature_sampling
instruction: This field lets the model know what task it needs to perform. Base models have been trained over a large set of varied instructions. You can give a simple and intuitive description of the task and the model will follow it, e.g., "Classify this movie review as positive or negative" or "Translate this sentence to Danish". Do not specify this if your dataset already prepends the instruction to the inputs field.
eval_dataset: Optional Cloud storage path to an evaluation dataset. If provided, inference will be performed on this dataset after training. The dataset format is jsonl. Each example in the dataset must contain a field `input_text` that contains the prompt.
project: Project used to run custom jobs. If not specified the project used to run the pipeline will be used.
accelerator_type: One of 'TPU' or 'GPU'. If 'TPU' is specified, tuning components run in europe-west4. Otherwise tuning components run in us-central1 on GPUs. Default is 'GPU'.
location: Location used to run custom jobs. If not specified the location used to run the pipeline will be used.
tensorboard_resource_id: Optional tensorboard resource id in format `projects/{project_number}/locations/{location}/tensorboards/{tensorboard_id}`. If provided, tensorboard metrics will be uploaded to this location.
Returns:
model_resource_name: Path to the model uploaded to the Model Registry. This will be an empty string if the model was not deployed.
endpoint_resource_name: Path the Online Prediction Endpoint. This will be an empty string if the model was not deployed.
"""
# fmt: on
id_columns = ['content']
task = 'summarization@001'
deploy_model = True
output_prediction_gcs_path_a = infer.infer_pipeline(
large_model_reference=large_model_a_reference,
prompt_dataset=preference_prompt_dataset,
prompt_sequence_length=prompt_sequence_length,
target_sequence_length=target_sequence_length,
sampling_strategy=sampling_strategy,
instruction=instruction,
project=project,
location=location,
accelerator_type=accelerator_type,
).set_display_name('Inferrer A')
output_prediction_gcs_path_b = infer.infer_pipeline(
large_model_reference=large_model_b_reference,
prompt_dataset=preference_prompt_dataset,
prompt_sequence_length=prompt_sequence_length,
target_sequence_length=target_sequence_length,
sampling_strategy=sampling_strategy,
instruction=instruction,
project=project,
location=location,
accelerator_type=accelerator_type,
).set_display_name('Inferrer B')
inference_output_uri = (
preference_data_formatter.format_preference_input_data(
model_a_inference_dir_uri=output_prediction_gcs_path_a.output,
model_b_inference_dir_uri=output_prediction_gcs_path_b.output,
instruction=instruction,
)
.set_display_name('Prepare AI Feedback Input')
.output
)
autosxs = online_evaluation_pairwise.online_evaluation_pairwise(
inference_output_uri=inference_output_uri,
id_columns=id_columns,
task=task,
).set_display_name('Build AI Feedback')
preference_dataset = (
preference_data_formatter.format_preference_data(
input_uri=autosxs.outputs['judgments_uri']
)
.set_display_name('Build Preference Dataset')
.output
)
rlhf_outputs = (
rlhf.rlhf_pipeline(
prompt_dataset=prompt_dataset,
preference_dataset=preference_dataset,
large_model_reference=large_model_reference,
model_display_name=model_display_name,
prompt_sequence_length=prompt_sequence_length,
target_sequence_length=target_sequence_length,
reward_model_train_steps=reward_model_train_steps,
reinforcement_learning_train_steps=reinforcement_learning_train_steps,
reward_model_learning_rate_multiplier=reward_model_learning_rate_multiplier,
reinforcement_learning_rate_multiplier=reinforcement_learning_rate_multiplier,
instruction=instruction,
deploy_model=deploy_model,
eval_dataset=eval_dataset,
kl_coeff=kl_coeff,
project=project,
location=location,
tensorboard_resource_id=tensorboard_resource_id,
accelerator_type=accelerator_type,
)
.set_display_name('Reinforcement Learning From AI Feedback')
.outputs
)
return PipelineOutput(
model_resource_name=rlhf_outputs['model_resource_name'],
endpoint_resource_name=rlhf_outputs['endpoint_resource_name'],
)