The Gemini API’s fine tuning support provides a mechanism for curating output when you have a small dataset of input/output examples. For more details, check out the Model tuning guide and tutorial.
Method: tunedModels.create
Creates a tuned model. Check intermediate tuning progress (if any) through the google.longrunning.Operations
service.
Access status and results through the Operations service. Example: GET /v1/tunedModels/az2mb0bpw6i/operations/000-111-222
Endpoint
post
https://generativelanguage.googleapis.com/v1beta/tunedModels
Query parameters
tunedModelId
string
Optional. The unique id for the tuned model if specified. This value should be up to 40 characters, the first character must be a letter, the last could be a letter or a number. The id must match the regular expression: [a-z]([a-z0-9-]{0,38}[a-z0-9])?
.
Request body
The request body contains an instance of TunedModel
.
displayName
string
Optional. The name to display for this model in user interfaces. The display name must be up to 40 characters including spaces.
description
string
Optional. A short description of this model.
Required. The tuning task that creates the tuned model.
source_model
. The model used as the starting point for tuning. source_model
can be only one of the following:Optional. TunedModel to use as the starting point for training the new model.
baseModel
string
Immutable. The name of the Model
to tune. Example: models/gemini-1.5-flash-001
temperature
number
Optional. Controls the randomness of the output.
Values can range over [0.0,1.0]
, inclusive. A value closer to 1.0
will produce responses that are more varied, while a value closer to 0.0
will typically result in less surprising responses from the model.
This value specifies default to be the one used by the base model while creating the model.
topP
number
Optional. For Nucleus sampling.
Nucleus sampling considers the smallest set of tokens whose probability sum is at least topP
.
This value specifies default to be the one used by the base model while creating the model.
topK
integer
Optional. For Top-k sampling.
Top-k sampling considers the set of topK
most probable tokens. This value specifies default to be used by the backend while making the call to the model.
This value specifies default to be the one used by the base model while creating the model.
Example request
Python
Response body
This resource represents a long-running operation that is the result of a network API call.
If successful, the response body contains data with the following structure:
name
string
The server-assigned name, which is only unique within the same service that originally returns it. If you use the default HTTP mapping, the name
should be a resource name ending with operations/{unique_id}
.
metadata
object
Service-specific metadata associated with the operation. It typically contains progress information and common metadata such as create time. Some services might not provide such metadata. Any method that returns a long-running operation should document the metadata type, if any.
An object containing fields of an arbitrary type. An additional field "@type"
contains a URI identifying the type. Example: { "id": 1234, "@type": "types.example.com/standard/id" }
.
done
boolean
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.
result
. The operation result, which can be either an error
or a valid response
. If done
== false
, neither error
nor response
is set. If done
== true
, exactly one of error
or response
can be set. Some services might not provide the result. result
can be only one of the following:The error result of the operation in case of failure or cancellation.
response
object
The normal, successful response of the operation. If the original method returns no data on success, such as Delete
, the response is google.protobuf.Empty
. If the original method is standard Get
/Create
/Update
, the response should be the resource. For other methods, the response should have the type XxxResponse
, where Xxx
is the original method name. For example, if the original method name is TakeSnapshot()
, the inferred response type is TakeSnapshotResponse
.
An object containing fields of an arbitrary type. An additional field "@type"
contains a URI identifying the type. Example: { "id": 1234, "@type": "types.example.com/standard/id" }
.
JSON representation |
---|
{ "name": string, "metadata": { "@type": string, field1: ..., ... }, "done": boolean, // Union field |
Method: tunedModels.generateContent
Generates a model response given an input GenerateContentRequest
. Refer to the text generation guide for detailed usage information. Input capabilities differ between models, including tuned models. Refer to the model guide and tuning guide for details.
Endpoint
post
https://generativelanguage.googleapis.com/v1beta/{model=tunedModels/*}:generateContent
Path parameters
model
string
Required. The name of the Model
to use for generating the completion.
Format: name=models/{model}
. It takes the form tunedModels/{tunedmodel}
.
Request body
The request body contains data with the following structure:
Optional. A list of Tools
the Model
may use to generate the next response.
A Tool
is a piece of code that enables the system to interact with external systems to perform an action, or set of actions, outside of knowledge and scope of the Model
. Supported Tool
s are Function
and codeExecution
. Refer to the Function calling and the Code execution guides to learn more.
Optional. Tool configuration for any Tool
specified in the request. Refer to the Function calling guide for a usage example.
Optional. A list of unique SafetySetting
instances for blocking unsafe content.
This will be enforced on the GenerateContentRequest.contents
and GenerateContentResponse.candidates
. There should not be more than one setting for each SafetyCategory
type. The API will block any contents and responses that fail to meet the thresholds set by these settings. This list overrides the default settings for each SafetyCategory
specified in the safetySettings. If there is no SafetySetting
for a given SafetyCategory
provided in the list, the API will use the default safety setting for that category. Harm categories HARM_CATEGORY_HATE_SPEECH, HARM_CATEGORY_SEXUALLY_EXPLICIT, HARM_CATEGORY_DANGEROUS_CONTENT, HARM_CATEGORY_HARASSMENT are supported. Refer to the guide for detailed information on available safety settings. Also refer to the Safety guidance to learn how to incorporate safety considerations in your AI applications.
Optional. Developer set system instruction(s). Currently, text only.
Optional. Configuration options for model generation and outputs.
cachedContent
string
Optional. The name of the content cached to use as context to serve the prediction. Format: cachedContents/{cachedContent}
Example request
Text
Python
Node.js
Go
Shell
Kotlin
Swift
Dart
Java
Image
Python
Node.js
Go
Shell
Kotlin
Swift
Dart
Java
Audio
Python
Node.js
Shell
Video
Python
Node.js
Go
Shell
Python
Shell
Chat
Python
Node.js
Go
Shell
Kotlin
Swift
Dart
Java
Cache
Python
Node.js
Tuned Model
Python
JSON Mode
Python
Node.js
Go
Shell
Kotlin
Swift
Dart
Java
Code execution
Python
Kotlin
Java
Function Calling
Python
Node.js
Shell
Kotlin
Swift
Dart
Java
Generation config
Python
Node.js
Go
Shell
Kotlin
Swift
Dart
Java
Safety Settings
Python
Node.js
Go
Shell
Kotlin
Swift
Dart
Java
System Instruction
Python
Node.js
Go
Shell
Kotlin
Swift
Dart
Java
Response body
If successful, the response body contains an instance of GenerateContentResponse
.
Method: tunedModels.get
Gets information about a specific TunedModel.
Endpoint
get
https://generativelanguage.googleapis.com/v1beta/{name=tunedModels/*}
Path parameters
name
string
Required. The resource name of the model.
Format: tunedModels/my-model-id
It takes the form tunedModels/{tunedmodel}
.
Request body
The request body must be empty.
Example request
Python
Response body
If successful, the response body contains an instance of TunedModel
.
Method: tunedModels.list
Lists created tuned models.
Endpoint
get
https://generativelanguage.googleapis.com/v1beta/tunedModels
Query parameters
pageSize
integer
Optional. The maximum number of TunedModels
to return (per page). The service may return fewer tuned models.
If unspecified, at most 10 tuned models will be returned. This method returns at most 1000 models per page, even if you pass a larger pageSize.
pageToken
string
Optional. A page token, received from a previous tunedModels.list
call.
Provide the pageToken
returned by one request as an argument to the next request to retrieve the next page.
When paginating, all other parameters provided to tunedModels.list
must match the call that provided the page token.
filter
string
Optional. A filter is a full text search over the tuned model's description and display name. By default, results will not include tuned models shared with everyone.
Additional operators: - owner:me - writers:me - readers:me - readers:everyone
Examples: "owner:me" returns all tuned models to which caller has owner role "readers:me" returns all tuned models to which caller has reader role "readers:everyone" returns all tuned models that are shared with everyone
Request body
The request body must be empty.
Example request
Python
Response body
Response from tunedModels.list
containing a paginated list of Models.
If successful, the response body contains data with the following structure:
The returned Models.
nextPageToken
string
A token, which can be sent as pageToken
to retrieve the next page.
If this field is omitted, there are no more pages.
JSON representation |
---|
{
"tunedModels": [
{
object ( |
Method: tunedModels.patch
Updates a tuned model.
Endpoint
patch
https://generativelanguage.googleapis.com/v1beta/{tunedModel.name=tunedModels/*}
PATCH https://generativelanguage.googleapis.com/v1beta/{tunedModel.name=tunedModels/*}
Path parameters
tunedModel.name
string
Output only. The tuned model name. A unique name will be generated on create. Example: tunedModels/az2mb0bpw6i
If displayName is set on create, the id portion of the name will be set by concatenating the words of the displayName with hyphens and adding a random portion for uniqueness.
Example:
- displayName =
Sentence Translator
- name =
tunedModels/sentence-translator-u3b7m
It takes the formtunedModels/{tunedmodel}
.
Query parameters
Required. The list of fields to update.
This is a comma-separated list of fully qualified names of fields. Example: "user.displayName,photo"
.
Request body
The request body contains an instance of TunedModel
.
displayName
string
Optional. The name to display for this model in user interfaces. The display name must be up to 40 characters including spaces.
description
string
Optional. A short description of this model.
Required. The tuning task that creates the tuned model.
source_model
. The model used as the starting point for tuning. source_model
can be only one of the following:Optional. TunedModel to use as the starting point for training the new model.
temperature
number
Optional. Controls the randomness of the output.
Values can range over [0.0,1.0]
, inclusive. A value closer to 1.0
will produce responses that are more varied, while a value closer to 0.0
will typically result in less surprising responses from the model.
This value specifies default to be the one used by the base model while creating the model.
topP
number
Optional. For Nucleus sampling.
Nucleus sampling considers the smallest set of tokens whose probability sum is at least topP
.
This value specifies default to be the one used by the base model while creating the model.
topK
integer
Optional. For Top-k sampling.
Top-k sampling considers the set of topK
most probable tokens. This value specifies default to be used by the backend while making the call to the model.
This value specifies default to be the one used by the base model while creating the model.
Response body
If successful, the response body contains an instance of TunedModel
.
Method: tunedModels.delete
Deletes a tuned model.
Endpoint
delete
https://generativelanguage.googleapis.com/v1beta/{name=tunedModels/*}
Path parameters
name
string
Required. The resource name of the model. Format: tunedModels/my-model-id
It takes the form tunedModels/{tunedmodel}
.
Request body
The request body must be empty.
Response body
If successful, the response body is empty.
REST Resource: tunedModels
- Resource: TunedModel
- TunedModelSource
- State
- TuningTask
- TuningSnapshot
- Dataset
- TuningExamples
- TuningExample
- Hyperparameters
- Methods
Resource: TunedModel
A fine-tuned model created using ModelService.CreateTunedModel.
name
string
Output only. The tuned model name. A unique name will be generated on create. Example: tunedModels/az2mb0bpw6i
If displayName is set on create, the id portion of the name will be set by concatenating the words of the displayName with hyphens and adding a random portion for uniqueness.
Example:
- displayName =
Sentence Translator
- name =
tunedModels/sentence-translator-u3b7m
displayName
string
Optional. The name to display for this model in user interfaces. The display name must be up to 40 characters including spaces.
description
string
Optional. A short description of this model.
Output only. The state of the tuned model.
Output only. The timestamp when this model was created.
A timestamp in RFC3339 UTC "Zulu" format, with nanosecond resolution and up to nine fractional digits. Examples: "2014-10-02T15:01:23Z"
and "2014-10-02T15:01:23.045123456Z"
.
Output only. The timestamp when this model was updated.
A timestamp in RFC3339 UTC "Zulu" format, with nanosecond resolution and up to nine fractional digits. Examples: "2014-10-02T15:01:23Z"
and "2014-10-02T15:01:23.045123456Z"
.
Required. The tuning task that creates the tuned model.
source_model
. The model used as the starting point for tuning. source_model
can be only one of the following:Optional. TunedModel to use as the starting point for training the new model.
baseModel
string
Immutable. The name of the Model
to tune. Example: models/gemini-1.5-flash-001
temperature
number
Optional. Controls the randomness of the output.
Values can range over [0.0,1.0]
, inclusive. A value closer to 1.0
will produce responses that are more varied, while a value closer to 0.0
will typically result in less surprising responses from the model.
This value specifies default to be the one used by the base model while creating the model.
topP
number
Optional. For Nucleus sampling.
Nucleus sampling considers the smallest set of tokens whose probability sum is at least topP
.
This value specifies default to be the one used by the base model while creating the model.
topK
integer
Optional. For Top-k sampling.
Top-k sampling considers the set of topK
most probable tokens. This value specifies default to be used by the backend while making the call to the model.
This value specifies default to be the one used by the base model while creating the model.
JSON representation |
---|
{ "name": string, "displayName": string, "description": string, "state": enum ( |
TunedModelSource
Tuned model as a source for training a new model.
tunedModel
string
Immutable. The name of the TunedModel
to use as the starting point for training the new model. Example: tunedModels/my-tuned-model
baseModel
string
Output only. The name of the base Model
this TunedModel
was tuned from. Example: models/gemini-1.5-flash-001
JSON representation |
---|
{ "tunedModel": string, "baseModel": string } |
State
The state of the tuned model.
Enums | |
---|---|
STATE_UNSPECIFIED |
The default value. This value is unused. |
CREATING |
The model is being created. |
ACTIVE |
The model is ready to be used. |
FAILED |
The model failed to be created. |
TuningTask
Tuning tasks that create tuned models.
Output only. The timestamp when tuning this model started.
A timestamp in RFC3339 UTC "Zulu" format, with nanosecond resolution and up to nine fractional digits. Examples: "2014-10-02T15:01:23Z"
and "2014-10-02T15:01:23.045123456Z"
.
Output only. The timestamp when tuning this model completed.
A timestamp in RFC3339 UTC "Zulu" format, with nanosecond resolution and up to nine fractional digits. Examples: "2014-10-02T15:01:23Z"
and "2014-10-02T15:01:23.045123456Z"
.
Output only. Metrics collected during tuning.
Required. Input only. Immutable. The model training data.
Immutable. Hyperparameters controlling the tuning process. If not provided, default values will be used.
JSON representation |
---|
{ "startTime": string, "completeTime": string, "snapshots": [ { object ( |
TuningSnapshot
Record for a single tuning step.
step
integer
Output only. The tuning step.
epoch
integer
Output only. The epoch this step was part of.
meanLoss
number
Output only. The mean loss of the training examples for this step.
Output only. The timestamp when this metric was computed.
A timestamp in RFC3339 UTC "Zulu" format, with nanosecond resolution and up to nine fractional digits. Examples: "2014-10-02T15:01:23Z"
and "2014-10-02T15:01:23.045123456Z"
.
JSON representation |
---|
{ "step": integer, "epoch": integer, "meanLoss": number, "computeTime": string } |
Dataset
Dataset for training or validation.
dataset
. Inline data or a reference to the data. dataset
can be only one of the following:Optional. Inline examples.
JSON representation |
---|
{ // Union field |
TuningExamples
A set of tuning examples. Can be training or validation data.
Required. The examples. Example input can be for text or discuss, but all examples in a set must be of the same type.
JSON representation |
---|
{
"examples": [
{
object ( |
TuningExample
A single example for tuning.
output
string
Required. The expected model output.
model_input
. The input to the model for this example. model_input
can be only one of the following:textInput
string
Optional. Text model input.
JSON representation |
---|
{ "output": string, // Union field |
Hyperparameters
Hyperparameters controlling the tuning process. Read more at https://ai.google.dev/docs/model_tuning_guidance
learning_rate_option
. Options for specifying learning rate during tuning. learning_rate_option
can be only one of the following:learningRate
number
Optional. Immutable. The learning rate hyperparameter for tuning. If not set, a default of 0.001 or 0.0002 will be calculated based on the number of training examples.
learningRateMultiplier
number
Optional. Immutable. The learning rate multiplier is used to calculate a final learningRate based on the default (recommended) value. Actual learning rate := learningRateMultiplier * default learning rate Default learning rate is dependent on base model and dataset size. If not set, a default of 1.0 will be used.
epochCount
integer
Immutable. The number of training epochs. An epoch is one pass through the training data. If not set, a default of 5 will be used.
batchSize
integer
Immutable. The batch size hyperparameter for tuning. If not set, a default of 4 or 16 will be used based on the number of training examples.
JSON representation |
---|
{ // Union field |