Foundation Models Tuning Service API, gRPC: TuningService.Tune

Статья создана

Обновлена 17 декабря 2024 г.

gRPC request
TuningRequest
WeightedDataset
TextToTextCompletionTuningParams
TuningTypeLora
TuningTypePromptTune
Scheduler
SchedulerLinear
SchedulerConstant
SchedulerCosine
Optimizer
OptimizerAdamw
TextClassificationMultilabelParams
Scheduler
Optimizer
TextClassificationMulticlassParams
Scheduler
Optimizer
yandex.cloud.operation.Operation
TuningMetadata
TuningResponse

gRPC request

rpc Tune (TuningRequest) returns (yandex.cloud.operation.Operation)

TuningRequest

{
  "base_model_uri": "string",
  "train_datasets": [
    {
      "dataset_id": "string",
      "weight": "double"
    }
  ],
  "validation_datasets": [
    {
      "dataset_id": "string",
      "weight": "double"
    }
  ],
  // Includes only one of the fields `text_to_text_completion`, `text_classification_multilabel`, `text_classification_multiclass`
  "text_to_text_completion": {
    "seed": "int64",
    "lr": "double",
    "n_samples": "int64",
    "additional_arguments": "string",
    // Includes only one of the fields `lora`, `prompt_tune`
    "lora": {
      "rank": "int64",
      "alpha": "double",
      "initialization": "string",
      "type": "string"
    },
    "prompt_tune": {
      "virtual_tokens": "int64"
    },
    // end of the list of possible fields
    "scheduler": {
      // Includes only one of the fields `linear`, `constant`, `cosine`
      "linear": {
        "min_lr": "double"
      },
      "constant": "SchedulerConstant",
      "cosine": {
        "min_lr": "double"
      },
      // end of the list of possible fields
      "warmup_ratio": "double"
    },
    "optimizer": {
      // Includes only one of the fields `adamw`
      "adamw": {
        "beta1": "double",
        "beta2": "double",
        "eps": "double",
        "weight_decay": "double"
      }
      // end of the list of possible fields
    }
  },
  "text_classification_multilabel": {
    "seed": "int64",
    "lr": "double",
    "n_samples": "int64",
    "additional_arguments": "string",
    // Includes only one of the fields `lora`, `prompt_tune`
    "lora": {
      "rank": "int64",
      "alpha": "double",
      "initialization": "string",
      "type": "string"
    },
    "prompt_tune": {
      "virtual_tokens": "int64"
    },
    // end of the list of possible fields
    "scheduler": {
      // Includes only one of the fields `linear`, `constant`, `cosine`
      "linear": {
        "min_lr": "double"
      },
      "constant": "SchedulerConstant",
      "cosine": {
        "min_lr": "double"
      },
      // end of the list of possible fields
      "warmup_ratio": "double"
    },
    "optimizer": {
      // Includes only one of the fields `adamw`
      "adamw": {
        "beta1": "double",
        "beta2": "double",
        "eps": "double",
        "weight_decay": "double"
      }
      // end of the list of possible fields
    }
  },
  "text_classification_multiclass": {
    "seed": "int64",
    "lr": "double",
    "n_samples": "int64",
    "additional_arguments": "string",
    // Includes only one of the fields `lora`, `prompt_tune`
    "lora": {
      "rank": "int64",
      "alpha": "double",
      "initialization": "string",
      "type": "string"
    },
    "prompt_tune": {
      "virtual_tokens": "int64"
    },
    // end of the list of possible fields
    "scheduler": {
      // Includes only one of the fields `linear`, `constant`, `cosine`
      "linear": {
        "min_lr": "double"
      },
      "constant": "SchedulerConstant",
      "cosine": {
        "min_lr": "double"
      },
      // end of the list of possible fields
      "warmup_ratio": "double"
    },
    "optimizer": {
      // Includes only one of the fields `adamw`
      "adamw": {
        "beta1": "double",
        "beta2": "double",
        "eps": "double",
        "weight_decay": "double"
      }
      // end of the list of possible fields
    }
  },
  // end of the list of possible fields
  "name": "string",
  "description": "string",
  "labels": "map<string, string>"
}

Field	Description
base_model_uri	string Required field. Format like a gpt://{folder_id}/yandex-gpt/latest
train_datasets[]	WeightedDataset
validation_datasets[]	WeightedDataset
text_to_text_completion	TextToTextCompletionTuningParams Includes only one of the fields `text_to_text_completion`, `text_classification_multilabel`, `text_classification_multiclass`.
text_classification_multilabel	TextClassificationMultilabelParams Includes only one of the fields `text_to_text_completion`, `text_classification_multilabel`, `text_classification_multiclass`.
text_classification_multiclass	TextClassificationMulticlassParams Includes only one of the fields `text_to_text_completion`, `text_classification_multilabel`, `text_classification_multiclass`.
name	string common params
description	string
labels	object (map<string, string>)

WeightedDataset

Field

Description

dataset_id

string

Required field.

weight

double

TextToTextCompletionTuningParams

Field	Description
seed	int64
lr	double
n_samples	int64
additional_arguments	string
lora	TuningTypeLora Includes only one of the fields `lora`, `prompt_tune`.
prompt_tune	TuningTypePromptTune Includes only one of the fields `lora`, `prompt_tune`.
scheduler	Scheduler
optimizer	Optimizer

TuningTypeLora

Field	Description
rank	int64
alpha	double
initialization	string
type	string

TuningTypePromptTune

Field	Description
virtual_tokens	int64

Scheduler

Field	Description
linear	SchedulerLinear Includes only one of the fields `linear`, `constant`, `cosine`.
constant	SchedulerConstant Includes only one of the fields `linear`, `constant`, `cosine`.
cosine	SchedulerCosine Includes only one of the fields `linear`, `constant`, `cosine`.
warmup_ratio	double

SchedulerLinear

Field	Description
min_lr	double

SchedulerConstant

Field	Description
Empty

SchedulerCosine

Field	Description
min_lr	double

Optimizer

Field

Description

adamw

OptimizerAdamw

Includes only one of the fields adamw.

OptimizerAdamw

Field	Description
beta1	double
beta2	double
eps	double
weight_decay	double

TextClassificationMultilabelParams

Field	Description
seed	int64
lr	double
n_samples	int64
additional_arguments	string
lora	TuningTypeLora Includes only one of the fields `lora`, `prompt_tune`.
prompt_tune	TuningTypePromptTune Includes only one of the fields `lora`, `prompt_tune`.
scheduler	Scheduler
optimizer	Optimizer

Scheduler

Field	Description
linear	SchedulerLinear Includes only one of the fields `linear`, `constant`, `cosine`.
constant	SchedulerConstant Includes only one of the fields `linear`, `constant`, `cosine`.
cosine	SchedulerCosine Includes only one of the fields `linear`, `constant`, `cosine`.
warmup_ratio	double

Optimizer

Field

Description

adamw

OptimizerAdamw

Includes only one of the fields adamw.

TextClassificationMulticlassParams

Field	Description
seed	int64
lr	double
n_samples	int64
additional_arguments	string
lora	TuningTypeLora Includes only one of the fields `lora`, `prompt_tune`.
prompt_tune	TuningTypePromptTune Includes only one of the fields `lora`, `prompt_tune`.
scheduler	Scheduler
optimizer	Optimizer

Scheduler

Field	Description
linear	SchedulerLinear Includes only one of the fields `linear`, `constant`, `cosine`.
constant	SchedulerConstant Includes only one of the fields `linear`, `constant`, `cosine`.
cosine	SchedulerCosine Includes only one of the fields `linear`, `constant`, `cosine`.
warmup_ratio	double

Optimizer

Field

Description

adamw

OptimizerAdamw

Includes only one of the fields adamw.

yandex.cloud.operation.Operation

{
  "id": "string",
  "description": "string",
  "created_at": "google.protobuf.Timestamp",
  "created_by": "string",
  "modified_at": "google.protobuf.Timestamp",
  "done": "bool",
  "metadata": {
    "tuning_task_id": "string",
    "status": "Status",
    "total_steps": "int64",
    "current_step": "int64"
  },
  // Includes only one of the fields `error`, `response`
  "error": "google.rpc.Status",
  "response": {
    "tuning_task_id": "string",
    "status": "Status",
    "target_model_uri": "string"
  }
  // end of the list of possible fields
}

An Operation resource. For more information, see Operation.

Field	Description
id	string ID of the operation.
description	string Description of the operation. 0-256 characters long.
created_at	google.protobuf.Timestamp Creation timestamp.
created_by	string ID of the user or service account who initiated the operation.
modified_at	google.protobuf.Timestamp The time when the Operation resource was last modified.
done	bool If the value is `false`, it means the operation is still in progress. If `true`, the operation is completed, and either `error` or `response` is available.
metadata	TuningMetadata Service-specific metadata associated with the operation. It typically contains the ID of the target resource that the operation is performed on. Any method that returns a long-running operation should document the metadata type, if any.
error	google.rpc.Status The error result of the operation in case of failure or cancellation. Includes only one of the fields `error`, `response`. The operation result. If `done == false` and there was no failure detected, neither `error` nor `response` is set. If `done == false` and there was a failure detected, `error` is set. If `done == true`, exactly one of `error` or `response` is set.
response	TuningResponse The normal response of the operation in case of success. If the original method returns no data on success, such as Delete, the response is google.protobuf.Empty. If the original method is the standard Create/Update, the response should be the target resource of the operation. Any method that returns a long-running operation should document the response type, if any. Includes only one of the fields `error`, `response`. The operation result. If `done == false` and there was no failure detected, neither `error` nor `response` is set. If `done == false` and there was a failure detected, `error` is set. If `done == true`, exactly one of `error` or `response` is set.

TuningMetadata

Field	Description
tuning_task_id	string
status	enum Status `STATUS_UNSPECIFIED` `CREATED` `PENDING` `IN_PROGRESS` `COMPLETED` `FAILED` `CANCELED`
total_steps	int64
current_step	int64

TuningResponse

Field	Description
tuning_task_id	string
status	enum Status `STATUS_UNSPECIFIED` `CREATED` `PENDING` `IN_PROGRESS` `COMPLETED` `FAILED` `CANCELED`
target_model_uri	string

Foundation Models Tuning Service API, gRPC: TuningService.Tune

gRPC requestgRPC request

TuningRequestTuningRequest

WeightedDatasetWeightedDataset

TextToTextCompletionTuningParamsTextToTextCompletionTuningParams

TuningTypeLoraTuningTypeLora

TuningTypePromptTuneTuningTypePromptTune

SchedulerScheduler

SchedulerLinearSchedulerLinear

SchedulerConstantSchedulerConstant

SchedulerCosineSchedulerCosine

OptimizerOptimizer

OptimizerAdamwOptimizerAdamw

TextClassificationMultilabelParamsTextClassificationMultilabelParams

SchedulerScheduler

OptimizerOptimizer

TextClassificationMulticlassParamsTextClassificationMulticlassParams

SchedulerScheduler

OptimizerOptimizer

yandex.cloud.operation.Operationyandex.cloud.operation.Operation

TuningMetadataTuningMetadata

TuningResponseTuningResponse

Была ли статья полезна?

gRPC request

TuningRequest

WeightedDataset

TextToTextCompletionTuningParams

TuningTypeLora

TuningTypePromptTune

Scheduler

SchedulerLinear

SchedulerConstant

SchedulerCosine

Optimizer

OptimizerAdamw

TextClassificationMultilabelParams

Scheduler

Optimizer

TextClassificationMulticlassParams

Scheduler

Optimizer

yandex.cloud.operation.Operation

TuningMetadata

TuningResponse