Add vmg templates, dataset_validation_util and update common_util (#3586)

* Add vmg templates, dataset_validation_util and update common_util * Add name to CODEOWNERS * Update common_util.py --------- Co-authored-by: Rayan Dasoriya <dasoriya@google.com>
2026-09-26 14:42:04 +00:00 · 2024-09-26 13:16:09 +00:00
parent 30c3e627a7
commit 0727e19520
14 changed files with 693 additions and 17 deletions
@@ -20,6 +20,7 @@
 /vertex_model_garden/model_oss/movinet @KCFindstr
 /vertex_model_garden/model_oss/data_converter @KCFindstr
 /vertex_model_garden/model_oss/peft @weigary
+/vertex_model_garden/model_oss/peft/templates @rayandasoriya
 /vertex_model_garden/model_oss/lm-evaluation-harness @kathyyu-google
 /vertex_model_garden/model_oss/tfvision @dstnluong-google
 /vertex_model_garden/model_oss/fvlm @minwoo33park
@@ -476,6 +476,7 @@ def get_resource_id(
    accelerator_type: str,
    is_for_training: bool,
    is_restricted_image: bool = False,
+    is_dynamic_workload_scheduler: bool = False,
 ) -> str:
  """Returns the resource id for a given accelerator type and the use case.

@@ -484,45 +485,61 @@ def get_resource_id(
    is_for_training: Whether the resource is used for training. Set false for
      serving use case.
    is_restricted_image: Whether the image is hosted in `vertex-ai-restricted`.
+    is_dynamic_workload_scheduler: Whether the resource is used with Dynamic
+      Workload Scheduler.

  Returns:
    The resource id.
  """
+  accelerator_suffix_map = {
+      "NVIDIA_TESLA_V100": "nvidia_v100_gpus",
+      "NVIDIA_L4": "nvidia_l4_gpus",
+      "NVIDIA_TESLA_A100": "nvidia_a100_gpus",
+      "NVIDIA_A100_80GB": "nvidia_a100_80gb_gpus",
+      "NVIDIA_H100_80GB": "nvidia_h100_gpus",
+      "NVIDIA_TESLA_T4": "nvidia_t4_gpus",
+      "TPU_V5e": "tpu_v5e",
+      "TPU_V3": "tpu_v3",
+  }
  default_training_accelerator_map = {
-      "NVIDIA_TESLA_V100": "custom_model_training_nvidia_v100_gpus",
-      "NVIDIA_L4": "custom_model_training_nvidia_l4_gpus",
-      "NVIDIA_TESLA_A100": "custom_model_training_nvidia_a100_gpus",
-      "NVIDIA_A100_80GB": "custom_model_training_nvidia_a100_80gb_gpus",
-      "NVIDIA_H100_80GB": "custom_model_training_nvidia_h100_gpus",
-      "NVIDIA_TESLA_T4": "custom_model_training_nvidia_t4_gpus",
-      "TPU_V5e": "custom_model_training_tpu_v5e",
-      "TPU_V3": "custom_model_training_tpu_v3",
+      key: f"custom_model_training_{accelerator_suffix_map[key]}"
+      for key in accelerator_suffix_map
+  }
+  dws_training_accelerator_map = {
+      key: f"custom_model_training_preemptible_{accelerator_suffix_map[key]}"
+      for key in accelerator_suffix_map
  }
  restricted_image_training_accelerator_map = {
      "NVIDIA_A100_80GB": "restricted_image_training_nvidia_a100_80gb_gpus",
  }
  serving_accelerator_map = {
-      "NVIDIA_TESLA_V100": "custom_model_serving_nvidia_v100_gpus",
-      "NVIDIA_L4": "custom_model_serving_nvidia_l4_gpus",
-      "NVIDIA_TESLA_A100": "custom_model_serving_nvidia_a100_gpus",
-      "NVIDIA_A100_80GB": "custom_model_serving_nvidia_a100_80gb_gpus",
-      "NVIDIA_H100_80GB": "custom_model_serving_nvidia_h100_gpus",
-      "NVIDIA_TESLA_T4": "custom_model_serving_nvidia_t4_gpus",
-      "TPU_V5e": "custom_model_serving_tpu_v5e",
+      key: f"custom_model_serving_{accelerator_suffix_map[key]}"
+      for key in accelerator_suffix_map
  }
+
  if is_for_training:
+    if is_restricted_image and is_dynamic_workload_scheduler:
+      raise ValueError(
+          "Dynamic Workload Scheduler does not work for restricted image"
+          " training."
+      )
    training_accelerator_map = (
        restricted_image_training_accelerator_map
        if is_restricted_image
        else default_training_accelerator_map
    )
    if accelerator_type in training_accelerator_map:
-      return training_accelerator_map[accelerator_type]
+      if is_dynamic_workload_scheduler:
+        return dws_training_accelerator_map[accelerator_type]
+      else:
+        return training_accelerator_map[accelerator_type]
    else:
      raise ValueError(
          f"Could not find accelerator type: {accelerator_type} for training."
      )
  else:
+    if is_dynamic_workload_scheduler:
+      raise ValueError("Dynamic Workload Scheduler does not work for serving.")
    if accelerator_type in serving_accelerator_map:
      return serving_accelerator_map[accelerator_type]
    else:
@@ -538,10 +555,14 @@ def check_quota(
    accelerator_count: int,
    is_for_training: bool,
    is_restricted_image: bool = False,
+    is_dynamic_workload_scheduler: bool = False,
 ):
  """Checks if the project and the region has the required quota."""
  resource_id = get_resource_id(
-      accelerator_type, is_for_training, is_restricted_image
+      accelerator_type,
+      is_for_training=is_for_training,
+      is_restricted_image=is_restricted_image,
+      is_dynamic_workload_scheduler=is_dynamic_workload_scheduler,
  )
  quota = get_quota(project_id, region, resource_id)
  quota_request_instruction = (
@@ -0,0 +1,486 @@
+"""Functions for dataset validation.
+
+This tool is used to validate the dataset against the given template.
+"""
+
+import json
+import multiprocessing
+import os
+import subprocess
+from typing import Any, Callable, Dict, Union
+from absl import logging
+import accelerate
+import datasets
+import transformers
+
+GCS_URI_PREFIX = "gs://"
+GCSFUSE_URI_PREFIX = "/gcs/"
+LOCAL_BASE_MODEL_DIR = "/tmp/base_model_dir"
+LOCAL_TEMPLATE_DIR = "/tmp/template_dir"
+_TEMPLATE_DIRNAME = "templates"
+_VERTEX_AI_SAMPLES_GITHUB_REPO_NAME = "vertex-ai-samples"
+_VERTEX_AI_SAMPLES_GITHUB_TEMPLATE_DIR = (
+    "community-content/vertex_model_garden/model_oss/peft/templates"
+)
+_DESCRIPTION_KEY = "description"
+_SOURCE_KEY = "source"
+_PROMPT_INPUT_KEY = "prompt_input"
+_PROMPT_NO_INPUT_KEY = "prompt_no_input"
+_RESPONSE_SEPARATOR = "response_separator"
+_INSTRUCTION_SEPARATOR = "instruction_separator"
+_CHAT_TEMPLATE_KEY = "chat_template"
+_KNOWN_KEYS = (
+    _DESCRIPTION_KEY,
+    _SOURCE_KEY,
+    _PROMPT_INPUT_KEY,
+    _PROMPT_NO_INPUT_KEY,
+    _RESPONSE_SEPARATOR,
+    _INSTRUCTION_SEPARATOR,
+    _CHAT_TEMPLATE_KEY,
+)
+
+
+def is_gcs_path(input_path: str) -> bool:
+  """Checks if the input path is a Google Cloud Storage (GCS) path.
+
+  Args:
+      input_path: The input path to be checked.
+
+  Returns:
+      True if the input path is a GCS path, False otherwise.
+  """
+  return input_path is not None and input_path.startswith(GCS_URI_PREFIX)
+
+
+def force_gcs_fuse_path(gcs_uri: str) -> str:
+  """Converts gs:// uris to their /gcs/ equivalents. No-op for other uris.
+
+  Args:
+    gcs_uri: The GCS URI to convert.
+
+  Returns:
+    The converted GCS URI.
+  """
+  if is_gcs_path(gcs_uri):
+    return GCSFUSE_URI_PREFIX + gcs_uri[len(GCS_URI_PREFIX) :]
+  else:
+    return gcs_uri
+
+
+def download_gcs_uri_to_local(
+    gcs_uri: str, destination_dir: str = LOCAL_BASE_MODEL_DIR
+) -> str:
+  """Downloads GCS URI to local.
+
+  If GCS URI is a directory, gs://some/folder is downloaded to
+  /destination_dir/folder. If GCS URI is a file, gs://some/file is downloaded to
+  /destination_dir/file.
+
+  Args:
+    gcs_uri: GCS URI to download.
+    destination_dir: Local directory directory.
+
+  Returns:
+    Local path to target folder/file.
+  """
+  target = os.path.join(
+      destination_dir,
+      os.path.basename(os.path.normpath(gcs_uri)),
+  )
+  if os.path.exists(target):
+    logging.info("File %s already exists.", target)
+    return target
+  if accelerate.PartialState().is_local_main_process:
+    logging.info(
+        "Downloading file(s) from %s to %s...", gcs_uri, destination_dir
+    )
+    if not os.path.exists(destination_dir):
+      os.mkdir(destination_dir)
+    subprocess.check_output([
+        "gsutil",
+        "-m",
+        "cp",
+        "-r",
+        gcs_uri,
+        destination_dir,
+    ])
+    logging.info("Downloaded file(s) from %s to %s.", gcs_uri, destination_dir)
+  # Make sure ALL processes process to next step after data downloading is done.
+  # It matters for the main process to wait for other processes as well.
+  accelerate.PartialState().wait_for_everyone()
+  return target
+
+
+def get_template(template_path: str) -> Dict[str, str]:
+  """Gets the template dictionary given the file path.
+
+  Args:
+    template_path: Path to the template file.
+
+  Returns:
+    A dictionary of the template.
+
+  Raises:
+    ValueError: If the template file does not exist or contains unknown keys.
+  """
+  if is_gcs_path(template_path):
+    template_path = force_gcs_fuse_path(template_path)
+  elif not os.path.isfile(template_path):
+    template_path = os.path.join(
+        os.path.dirname(__file__),
+        _TEMPLATE_DIRNAME,
+        template_path + ".json",
+    )
+  if not os.path.isfile(template_path):
+    raise ValueError(f"Template file {template_path} does not exist.")
+  with open(template_path, "r") as f:
+    template_json: dict[str, str] = json.load(f)
+    for key in template_json:
+      if key not in _KNOWN_KEYS:
+        raise ValueError(f"Unknown key {key} in template {template_path}.")
+    return template_json
+
+
+def get_response_separator(template_json: Dict[str, str]) -> Union[str, None]:
+  return template_json.get(_RESPONSE_SEPARATOR, None)
+
+
+def get_instruction_separator(
+    template_json: Dict[str, str],
+) -> Union[str, None]:
+  return template_json.get(_INSTRUCTION_SEPARATOR, None)
+
+
+def _format_template_fn(
+    template: str,
+    input_column: str,
+    tokenizer: transformers.PreTrainedTokenizer | None = None,
+) -> Callable[[Dict[str, str]], Dict[str, str]]:
+  """Formats a dataset example according to a template.
+
+  Args:
+    template: Name of the JSON template file under `templates/` or GCS path to
+      the template file.
+    input_column: The input column in the dataset to be used or updated by the
+      template. If it does not exist, the template's `prompt_no_input` will be
+      used, and the input_column will be created.
+    tokenizer: The tokenizer to use for chat_template templates.
+
+  Returns:
+    A function that formats data according to the template.
+  """
+  template_json = get_template(template)
+
+  if _CHAT_TEMPLATE_KEY not in template_json:
+
+    def format_fn(example: Dict[str, str]) -> Dict[str, str]:
+      format_dict = {key: value for key, value in example.items()}
+      format_str = (
+          template_json[_PROMPT_INPUT_KEY]
+          if format_dict.get(input_column)
+          else template_json[_PROMPT_NO_INPUT_KEY]
+      )
+      return {input_column: format_str.format(**format_dict)}
+
+    return format_fn
+  elif (
+      _PROMPT_INPUT_KEY in template_json
+      or _PROMPT_NO_INPUT_KEY in template_json
+  ):
+    raise ValueError(
+        "chat_template templates do not support input/no_input templates."
+    )
+  else:
+    if tokenizer is None:
+      raise ValueError("A tokenizer is required for chat_template templates.")
+    # Assign HuggingFace jinja template.
+    tokenizer.chat_template = template_json[_CHAT_TEMPLATE_KEY]
+    return lambda example: {
+        input_column: tokenizer.apply_chat_template(
+            example[input_column], tokenize=False, add_generation_prompt=False
+        )
+    }
+
+
+def _format_template_fn_notebook(
+    template: str,
+    input_column: str,
+    tokenizer: transformers.PreTrainedTokenizer | None = None,
+) -> Callable[[Dict[str, str]], Dict[str, str]]:
+  """Formats a dataset example according to a template.
+
+  Args:
+    template: Name of the JSON template file under `templates/` or GCS path to
+      the template file.
+    input_column: The input column in the dataset to be used or updated by the
+      template. If it does not exist, the template's `prompt_no_input` will be
+      used, and the input_column will be created.
+    tokenizer: The tokenizer to use for chat_template templates.
+
+  Returns:
+    A function that formats data according to the template.
+  """
+  template_json = get_template(template)
+
+  if _CHAT_TEMPLATE_KEY not in template_json:
+
+    def format_fn(example: Dict[str, str]) -> Dict[str, str]:
+      format_dict = {key: value for key, value in example.items()}
+      if format_dict.get(input_column):
+        format_str = template_json[_PROMPT_INPUT_KEY]
+      elif _PROMPT_NO_INPUT_KEY in template_json:
+        format_str = template_json[_PROMPT_NO_INPUT_KEY]
+      else:
+        raise KeyError(
+            f"The template {os.path.basename(template)} does not contain"
+            f" {_PROMPT_INPUT_KEY} or {_PROMPT_NO_INPUT_KEY} key."
+        )
+      try:
+        return {input_column: format_str.format(**format_dict)}
+      except KeyError as e:
+        raise KeyError(
+            f"The template {os.path.basename(template)} contains a key {e} in"
+            f" {_PROMPT_INPUT_KEY} or {_PROMPT_NO_INPUT_KEY} that does not"
+            " exist in the dataset example. The dataset example looks like"
+            f" {format_dict}."
+        ) from e
+
+    return format_fn
+  elif (
+      _PROMPT_INPUT_KEY in template_json
+      or _PROMPT_NO_INPUT_KEY in template_json
+  ):
+    raise ValueError(
+        f"chat_template templates do not support {_PROMPT_INPUT_KEY} or"
+        f" {_PROMPT_NO_INPUT_KEY} templates."
+    )
+  else:
+    if tokenizer is None:
+      raise ValueError("A tokenizer is required for chat_template templates.")
+    # Assign HuggingFace jinja template.
+    tokenizer.chat_template = template_json[_CHAT_TEMPLATE_KEY]
+
+    def format_fn(example: Dict[str, str]) -> Dict[str, str]:
+      try:
+        return {
+            input_column: tokenizer.apply_chat_template(
+                example[input_column],
+                tokenize=False,
+                add_generation_prompt=False,
+            )
+        }
+      except KeyError as e:
+        raise KeyError(
+            f"The template {os.path.basename(template)} contains a key {e} in"
+            f" {_CHAT_TEMPLATE_KEY} that does not exist in the dataset example."
+        ) from e
+
+    return format_fn
+
+
+def _get_split_string(
+    split: str,
+    dataset_percent: int | None = None,
+    dataset_k_rows: int | None = None,
+) -> str:
+  """Gets the formatted split string for the dataset.
+
+  This is used to format the split string as per
+  https://huggingface.co/docs/datasets/v2.21.0/loading#slice-splits. Also, this
+  function will only be used to load the partial dataset for validating the
+  dataset against the template.
+
+  Args:
+    split: Split of the dataset.
+    dataset_percent: The percentage of the dataset to load.
+    dataset_k_rows: The top k sequences to load from the dataset.
+
+  Returns:
+    A formatted split string.
+  """
+  # Validate the dataset_percent and dataset_k_rows values.
+  if dataset_percent and dataset_k_rows:
+    raise ValueError(
+        "You can set either validate_percentage_of_dataset or"
+        " validate_k_rows_of_dataset, but not both."
+    )
+
+  if dataset_percent:
+    logging.info("Loading %d percent of the dataset...", dataset_percent)
+    return f"{split}[:{dataset_percent}%]"
+
+  if dataset_k_rows:
+    logging.info("Loading top %d rows of the dataset...", dataset_k_rows)
+    return f"{split}[:{dataset_k_rows}]"
+
+  return split
+
+
+def _github_template_path(template: str) -> str:
+  """Generates the path to the template in the Vertex AI Samples GitHub repo.
+
+  Args:
+    template: Name of the template.
+
+  Returns:
+    The path to the template in the Vertex AI Samples GitHub repo.
+  """
+  # vertex-ai-samples directory may lie under separate directory depending on
+  # the scratch_dir parameter in the notebook execution environment.
+  vertex_ai_samples_abs_path = os.getcwd().split(
+      _VERTEX_AI_SAMPLES_GITHUB_REPO_NAME
+  )[0]
+  return os.path.join(
+      vertex_ai_samples_abs_path,
+      _VERTEX_AI_SAMPLES_GITHUB_REPO_NAME,
+      _VERTEX_AI_SAMPLES_GITHUB_TEMPLATE_DIR,
+      template + ".json",
+  )
+
+
+def _get_dataset(
+    dataset_name: str,
+    split: str,
+    num_proc: int | None = None,
+) -> datasets.DatasetDict:
+  """Gets a dataset.
+
+  Args:
+    dataset_name: Name of the dataset or path to a custom dataset.
+    split: Split of the dataset.
+    num_proc: Number of processors to use.
+
+  Returns:
+    A dataset.
+  """
+  dataset_name = force_gcs_fuse_path(dataset_name)
+  if os.path.isfile(dataset_name):
+    # Custom dataset.
+    return datasets.load_dataset(
+        "json",
+        data_files=[dataset_name],
+        split=split,
+        num_proc=num_proc,
+    )
+  # HF dataset.
+  return datasets.load_dataset(dataset_name, split=split, num_proc=num_proc)
+
+
+def load_dataset_with_template(
+    dataset_name: str,
+    split: str,
+    input_column: str,
+    template: str = None,
+    tokenizer: transformers.PreTrainedTokenizer | None = None,
+) -> Any:
+  """Loads dataset with templates.
+
+  Args:
+    dataset_name: Name of the dataset or path to a custom dataset.
+    split: Split of the dataset.
+    input_column: The input column in the dataset to be used or updaded by the
+      template. If it does not exist, the template's `prompt_no_input` will be
+      used, and the input_column will be created.
+    template: Name of the JSON template file under `templates/` or GCS path to
+      the template file.
+    tokenizer: The tokenizer to use for chat_template templates.
+
+  Returns:
+    A dataset compatible with the template.
+  """
+  dataset = _get_dataset(dataset_name, split=split)
+  if template:
+    dataset = dataset.map(
+        _format_template_fn(
+            template,
+            input_column=input_column,
+            tokenizer=tokenizer,
+        )
+    )
+
+  return dataset
+
+
+def validate_dataset_with_template(
+    dataset_name: str,
+    split: str,
+    input_column: str,
+    template: str,
+    tokenizer: transformers.PreTrainedTokenizer | None = None,
+    use_multiprocessing: bool = False,
+    validate_percentage_of_dataset: int | None = None,
+    validate_k_rows_of_dataset: int | None = None,
+) -> Any:
+  """Validates dataset with templates.
+
+  This function will be used to load the dataset and validate it against the
+  template. In case of validation, we also allow the users to load the dataset
+  partially by allowing them to read x% or top k rows of the dataset. To
+  validate the dataset, the template file must be available in the GCS bucket
+  and the dataset must be available either in the GCS bucket or Hugging Face.
+
+  Args:
+    dataset_name: Name of the dataset or path to a custom dataset.
+    split: Split of the dataset.
+    input_column: The input column in the dataset to be used or updaded by the
+      template. If it does not exist, the template's `prompt_no_input` will be
+      used, and the input_column will be created.
+    template: Name of the JSON template file under `templates/` or GCS path to
+      the template file.
+    tokenizer: The tokenizer to use for chat_template templates.
+    use_multiprocessing: If True, it will use multiprocessing to load the
+      dataset.
+    validate_percentage_of_dataset: The percentage of the dataset to load.
+    validate_k_rows_of_dataset: The top k sequences to load from the dataset.
+
+  Returns:
+    None if the validation is successful, otherwise returns the error message.
+  """
+  if not template:
+    raise ValueError("template is required for validate_dataset.")
+
+  if not dataset_name:
+    raise ValueError("dataset_name is empty.")
+
+  if not split:
+    raise ValueError("split is empty.")
+
+  split = _get_split_string(
+      split,
+      validate_percentage_of_dataset,
+      validate_k_rows_of_dataset,
+  )
+
+  num_proc = multiprocessing.cpu_count() if use_multiprocessing else 1
+
+  # gcsfuse cannot be used from the notebook runtime env. Hence, we have
+  # to download dataset and template from gcs to local.
+  if is_gcs_path(dataset_name):
+    dataset_name = download_gcs_uri_to_local(dataset_name, LOCAL_BASE_MODEL_DIR)
+
+  if is_gcs_path(template):
+    template_path = download_gcs_uri_to_local(template, LOCAL_TEMPLATE_DIR)
+  elif os.path.isfile(_github_template_path(template)):
+    template_path = _github_template_path(template)
+  else:
+    raise ValueError(
+        f"Template file {template} does not exist. To validate the"
+        " dataset, please provide a valid GCS path for the template or a valid"
+        " template name from"
+        f" https://github.com/GoogleCloudPlatform/{_VERTEX_AI_SAMPLES_GITHUB_REPO_NAME}/tree/main/{_VERTEX_AI_SAMPLES_GITHUB_TEMPLATE_DIR}."
+    )
+
+  _get_dataset(dataset_name, split, num_proc).map(
+      _format_template_fn_notebook(
+          template_path,
+          input_column=input_column,
+          tokenizer=tokenizer,
+      )
+  )
+
+  print(
+      "Dataset {} is compatible with the {} template.".format(
+          os.path.basename(dataset_name), os.path.basename(template)
+      )
+  )
+
@@ -0,0 +1,98 @@
+# Vertex Model Garden Training Dataset Template
+
+## Overview
+
+Vertex Model Garden training provides templates for streamlined preprocessing of
+datasets. Although datasets often have intricate structures, the supported LLM
+models accept only flat strings. A template facilitates parsing a dataset and
+preprocessing it to be compatible with the model.
+
+When fine-tuning a pretrained model, it is advisable to maintain the same format
+as the original training data. A template helps replicate the format, ensuring
+consistency and potentially enhancing the fine-tuning process.
+
+Both multi-turn messages and single instruction-response pairs are supported.
+Multi-turn messages are accommodated using a more general `chat_template` field,
+whereas simple instruction-response pair datasets are supported through the
+`prompt_input` field.
+
+A template is a JSON file consisting of string key-value pairs. Refer to the
+following for the definitions of the supported fields.
+
+## Template field documentation
+
+**description**: An explanation of the template.
+
+**source**: Information about the origin of the template.
+
+**chat_template**: A
+[jinja template](https://jinja.palletsprojects.com/en/3.1.x/templates/) that can
+be used to parse a chat dataset. This is the same format as
+[HF chat templates](https://huggingface.co/docs/transformers/main/en/chat_templating).
+To create a chat_template, use the `messages` variable to be filled with the
+sample. The flag `--instruct_column_in_dataset` identifies which column will be
+passed to the `messages` variable in the chat_template. This field is mutually
+exclusive with `prompt_input` and `prompt_no_input`.
+
+**prompt_input**: A string template that is used when value for the input column
+exists in the sample. It should be able to be formatted with the
+[str.format](https://docs.python.org/3/library/stdtypes.html#str.format) method.
+The input column is specified with the flag `--instruct_column_in_dataset`. Used
+for instruction dataset. This field is mutually exclusive with `chat_template`.
+
+**prompt_no_input**: A string template that is used when value for the input
+column does not exist in the sample. It should be able to be formatted with the
+[str.format](https://docs.python.org/3/library/stdtypes.html#str.format) method.
+The input column is specified with the flag `--instruct_column_in_dataset`. Used
+for instruction dataset. This field is mutually exclusive with `chat_template`.
+
+**instruction_separator**: A unique string used to indicate the start of the
+instructions. If not specified, every token after response_separator will be
+treated as a response, and every token before the first response_separator will
+be treated as instruction.
+
+**response_separator**: A unique string used to indicate the start of the
+response. This field is required if `--completion_only` flag is set to `True`.
+
+## Example templates
+
+-   See the list of all supported templates [here](https://github.com/GoogleCloudPlatform/vertex-ai-samples/tree/main/community-content/vertex_model_garden/model_oss/peft/train/vmg/templates).
+-   For an example with `chat_template` see the JSON template below.
+
+```
+{
+  "description": "Chat template used by Llama 3.",
+  "source": "https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct/blob/a5a71a7527eac1d651bb145436c72026887fb68e/tokenizer_config.json#L2053",
+  "chat_template": "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>assistant<|end_header_id|>\n\n"
+}
+```
+
+-   For an example with `prompt_input` see the JSON template below. In this case
+    the flag `--instruct_column_in_dataset=text` should be set, and there must
+    be a column named `text` in the dataset.
+
+```
+{
+  "description": "Template for openassistant-guanaco dataset.",
+  "source": "https://huggingface.co/datasets/timdettmers/openassistant-guanaco",
+  "prompt_input": "{text}",
+  "instruction_separator": "### Human:",
+  "response_separator": "### Assistant:"
+}
+```
+
+-   For an example with `prompt_no_input` see the JSON template below. In this
+    case the flag `--instruct_column_in_dataset=input` should be set, and there
+    must be columns named `input` and `instruction` in the dataset.
+
+```
+{
+  "description": "Template used by Alpaca-LoRA.",
+  "source": "https://github.com/tloen/alpaca-lora/blob/main/templates/alpaca.json",
+  "prompt_input": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n",
+  "prompt_no_input": "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response:\n",
+  "response_separator": "### Response:"
+}
+```
@@ -0,0 +1,7 @@
+{
+  "description": "Template used by Alpaca-LoRA.",
+  "source": "https://github.com/tloen/alpaca-lora/blob/main/templates/alpaca.json",
+  "prompt_input": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n",
+  "prompt_no_input": "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response:\n",
+  "response_separator": "### Response:"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "A shorter template to experiment with.",
+  "source": "https://github.com/tloen/alpaca-lora/blob/main/templates/alpaca_short.json",
+  "prompt_input": "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n",
+  "prompt_no_input": "### Instruction:\n{instruction}\n\n### Response:\n",
+  "response_separator": "### Response:"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Chat template used by Gemma. 'assistant' role is replaced by 'model'",
+  "source": "https://huggingface.co/google/gemma-1.1-2b-it/blob/bf4924f313df5166dee1467161e886e55f2eb4d4/tokenizer_config.json#L1507",
+  "chat_template": "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\n' + message['content'] | trim + '<end_of_turn>\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\n'}}{% endif %}",
+  "instruction_separator": "<start_of_turn>user\n",
+  "response_separator": "<start_of_turn>model\n"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Template used by Llama 3, accepting text-bison format.",
+  "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/models/tune-text-models-supervised#dataset-format",
+  "prompt_input": "\n\n<|start_header_id|>user<|end_header_id|>\n\n{input_text}<|eot_id|>\n\n<|start_header_id|>assistant<|end_header_id|>\n\n{output_text}<|eot_id|>",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>assistant<|end_header_id|>\n\n"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Chat template used by Llama 3.",
+  "source": "https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct/blob/a5a71a7527eac1d651bb145436c72026887fb68e/tokenizer_config.json#L2053",
+  "chat_template": "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>assistant<|end_header_id|>\n\n"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Chat template used by Mistral.",
+  "source": "https://github.com/OpenAccess-AI-Collective/axolotl/blob/main/src/axolotl/utils/chat_templates.py",
+  "chat_template": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",
+  "instruction_separator": "[INST]",
+  "response_separator": "[/INST]"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Template used by openai chat.",
+  "source": "https://platform.openai.com/docs/api-reference/fine-tuning/chat-input",
+  "chat_template": "{% set loop_messages = messages %}{% set content = '' %}{% for message in loop_messages %}{% set content = content ~ '\n\n<|start_header_id|>' ~ message.role ~ '<|end_header_id|>\n\n' %}{% if message.content is string %}{% set content = content ~ message.content|trim ~ '<|eot_id|>' %}{% else %}{% set content = content ~ message.content|join(' ', attribute='text')|trim ~ '<|eot_id|>' %}{% endif %}{% if loop.index0 == 0 %}{% set content = bos_token ~ content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>assistant<|end_header_id|>\n\n"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Template used by openai completion.",
+  "source": "https://platform.openai.com/docs/api-reference/fine-tuning/completions-input",
+  "prompt_input": "\n\n<|start_header_id|>user<|end_header_id|>\n\n{prompt}<|eot_id|>\n\n<|start_header_id|>assistant<|end_header_id|>\n\n{completion}<|eot_id|>",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>assistant<|end_header_id|>\n\n"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Template for openassistant-guanaco dataset.",
+  "source": "https://huggingface.co/datasets/timdettmers/openassistant-guanaco",
+  "prompt_input": "{text}",
+  "instruction_separator": "### Human:",
+  "response_separator": "### Assistant:"
+}
@@ -0,0 +1,7 @@
+{
+  "description": "Template used for chat based models.",
+  "source": "https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct/blob/a5a71a7527eac1d651bb145436c72026887fb68e/tokenizer_config.json#L2053",
+  "chat_template": "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>model<|end_header_id|>\n\n' }}{% endif %}",
+  "instruction_separator": "<|start_header_id|>user<|end_header_id|>\n\n",
+  "response_separator": "<|start_header_id|>model<|end_header_id|>\n\n"
+}