Signature
Method signature onclient.trainings:
replaceConfiguration(args: {
trainingId: string;
configuration: TrainingConfigurationInput;
}): Promise<Training>;
Parameters
string
required
Training draft ID with prefix
trg..TrainingConfigurationInput
required
Method-specific Training draft input.
Show method: "sft"
Show method: "sft"
"sft"
required
Training method;
sft learns a LoRA adapter from prompt and target examples.TrainingModelSource
required
The exact Base Model or Model Version used as the source.
Show type: "base_model"
Show type: "base_model"
string
required
The dataset slice to train on. It must freeze a split field. Training uses its train parts and validation uses its validate parts at the slice cutoff. Rows marked as test by any split field are excluded, except validation rows in a part with both validate and test roles in the selected split field.
string
The split field to use, identified by field ID. Required when the slice freezes more than one split field.
string[]
The train parts to learn on. Defaults to every part with the train role.
8 | 16 | 32
default:"16"
LoRA adapter rank: 8, 16, or 32. Defaults to 16. Alpha is twice the rank.
number
default:"3"
Passes over the training selection, 1 through 5. The cost estimate grows linearly with epochs.
number
default:"0.00015"
Optimizer learning rate, 0.00001 through 0.001. The default is 0.00015.
8 | 16 | 32 | 64
default:"8"
Sequences per optimizer step across all GPUs: 8, 16, 32, or 64. The default is 8.
number
default:"12288"
Token budget per training example, prompt and target together: 64 through 16384. The cost estimate is blocked when the longest example exceeds it. Defaults: 12288 for SFT, 3072 for gisting.
Show method: "gisting"
Show method: "gisting"
"gisting"
required
Training method;
gisting learns tokens that replace a reusable prompt.TrainingModelSource
required
The exact Base Model or Model Version used as the source.
Show type: "base_model"
Show type: "base_model"
string
required
The dataset slice to train on. It must freeze a split field. Training uses its train parts and validation uses its validate parts at the slice cutoff. Rows marked as test by any split field are excluded, except validation rows in a part with both validate and test roles in the selected split field.
string
The split field to use, identified by field ID. Required when the slice freezes more than one split field.
string[]
The train parts to learn on. Defaults to every part with the train role.
string
required
Project file ID with prefix
ppf. containing the finalized UTF-8 prompt to compress.8 | 32 | 64 | "1/4" | "1/8" | "1/16" | "1/32"
default:"1/8"
Number of learned tokens that replace the prompt: 8, 32, 64, or a fraction of the prompt length: “1/4”, “1/8”, “1/16”, or “1/32”. Fractions are rounded up to a whole token and limited to 8 through 256 tokens. Defaults to “1/8”.
number
default:"3"
Passes over the training selection, 1 through 5. The cost estimate grows linearly with epochs.
number
default:"0.03"
Optimizer learning rate for the soft tokens, 0.001 through 0.3. The default is 0.03.
8 | 16 | 32 | 64
default:"8"
Sequences per optimizer step across all GPUs: 8, 16, 32, or 64. The default is 8.
number
default:"3072"
Token budget per training example, prompt and target together: 64 through 16384. The cost estimate is blocked when the longest example exceeds it. Defaults: 12288 for SFT, 3072 for gisting.
Returns
Promise<Training>
A project-owned mutable Training draft. The output Model is bound on the first Training Run.
TrainingId
required
Training draft ID with prefix
trg..ProjectId
required
Project ID with prefix
pp..ModelId | null
required
Output model ID with prefix
mdl., or null before the first training run.string
required
Name of the training draft.
TrainingConfiguration
required
A normalized method-specific Training draft.
Show method: "sft"
Show method: "sft"
"sft"
required
Training method;
sft learns a LoRA adapter from prompt and target examples.TrainingModelSource
required
The exact Base Model or Model Version used as the source.
Show type: "base_model"
Show type: "base_model"
string
The dataset slice to train on. It must freeze a split field. Training uses its train parts and validation uses its validate parts at the slice cutoff. Rows marked as test by any split field are excluded, except validation rows in a part with both validate and test roles in the selected split field.
string
The split field to use, identified by field ID. Required when the slice freezes more than one split field.
string[]
The train parts to learn on. Defaults to every part with the train role.
8 | 16 | 32
required
LoRA adapter rank: 8, 16, or 32. Defaults to 16. Alpha is twice the rank.
number
required
Passes over the training selection, 1 through 5. The cost estimate grows linearly with epochs.
number
required
Optimizer learning rate, 0.00001 through 0.001. The default is 0.00015.
8 | 16 | 32 | 64
required
Sequences per optimizer step across all GPUs: 8, 16, 32, or 64. The default is 8.
number
required
Token budget per training example, prompt and target together: 64 through 16384. The cost estimate is blocked when the longest example exceeds it. Defaults: 12288 for SFT, 3072 for gisting.
TrainingDatasetSelection
required
A Dataset and normalized filter selected for one Training input purpose.
Show Fields
Show Fields
string
required
Dataset ID with prefix
ds. containing the selected examples.TrainingDatasetItemFilter
required
Conditions selecting dataset rows; omitted conditions do not restrict the selection.See the HTTP endpoint for the complete schema.
TrainingDatasetSelection
A Dataset and normalized filter selected for one Training input purpose.
Show Fields
Show Fields
string
required
Dataset ID with prefix
ds. containing the selected examples.TrainingDatasetItemFilter
required
Conditions selecting dataset rows; omitted conditions do not restrict the selection.See the HTTP endpoint for the complete schema.
Show method: "gisting"
Show method: "gisting"
"gisting"
required
Training method;
gisting learns tokens that replace a reusable prompt.TrainingModelSource
required
The exact Base Model or Model Version used as the source.
Show type: "base_model"
Show type: "base_model"
string
The dataset slice to train on. It must freeze a split field. Training uses its train parts and validation uses its validate parts at the slice cutoff. Rows marked as test by any split field are excluded, except validation rows in a part with both validate and test roles in the selected split field.
string
The split field to use, identified by field ID. Required when the slice freezes more than one split field.
string[]
The train parts to learn on. Defaults to every part with the train role.
string
required
Project file ID with prefix
ppf. containing the finalized UTF-8 prompt to compress.8 | 32 | 64 | "1/4" | "1/8" | "1/16" | "1/32"
required
Number of learned tokens that replace the prompt: 8, 32, 64, or a fraction of the prompt length: “1/4”, “1/8”, “1/16”, or “1/32”. Fractions are rounded up to a whole token and limited to 8 through 256 tokens. Defaults to “1/8”.
number
required
Passes over the training selection, 1 through 5. The cost estimate grows linearly with epochs.
number
required
Optimizer learning rate for the soft tokens, 0.001 through 0.3. The default is 0.03.
8 | 16 | 32 | 64
required
Sequences per optimizer step across all GPUs: 8, 16, 32, or 64. The default is 8.
number
required
Token budget per training example, prompt and target together: 64 through 16384. The cost estimate is blocked when the longest example exceeds it. Defaults: 12288 for SFT, 3072 for gisting.
TrainingDatasetSelection
required
A Dataset and normalized filter selected for one Training input purpose.
Show Fields
Show Fields
string
required
Dataset ID with prefix
ds. containing the selected examples.TrainingDatasetItemFilter
required
Conditions selecting dataset rows; omitted conditions do not restrict the selection.See the HTTP endpoint for the complete schema.
TrainingDatasetSelection
A Dataset and normalized filter selected for one Training input purpose.
Show Fields
Show Fields
string
required
Dataset ID with prefix
ds. containing the selected examples.TrainingDatasetItemFilter
required
Conditions selecting dataset rows; omitted conditions do not restrict the selection.See the HTTP endpoint for the complete schema.
string | null
required
Time the training was archived, as an ISO 8601 timestamp, or null while it is active.
string
required
Time the training was created, as an ISO 8601 timestamp.
string
required
Time the training was last updated, as an ISO 8601 timestamp.
Errors
HTTP errors throwPotatoApiError. The endpoint documents these error responses:
| Status | Description |
|---|---|
| 404 | Training, Dataset, Dataset Slice, Model Version, or gisting prompt file not found |
| 409 | Archived Training cannot be updated |
| 429 | Rate limit exceeded |
PUT /trainings/{trainingId}/configuration.
import { createClient } from "@toloka-platform/sdk";
const client = createClient({
baseUrl: "https://platform.toloka.ai",
apiKey: process.env.TOLOKA_API_KEY!,
});
const training = await client.trainings.replaceConfiguration({
trainingId: "trg.019a0000-0000-7000-8000-000000000001",
configuration: {
method: "sft",
source: {
type: "base_model",
baseModelId: "qwen3-4b"
},
sliceId: "dsl.019a0000-0000-7000-8000-000000000001"
}
});
console.log(training);
Nested values in this example are abbreviated. The HTTP endpoint has the complete schema.
{
"id": "trg.019a0000-0000-7000-8000-000000000001",
"projectId": "pp.019a0000-0000-7000-8000-000000000001",
"modelId": null,
"name": "<string>",
"configuration": {
"method": "sft",
"source": {
"type": "base_model",
"baseModelId": "qwen3-4b"
},
"loraRank": 8,
"epochs": 1,
"learningRate": 0.00001,
"batchSize": 8,
"maxSequenceLength": 64,
"training": {
"datasetId": "ds.019a0000-0000-7000-8000-000000000001",
"itemFilter": {
"fieldValues": [
{}
],
"fieldSources": [
{}
],
"missingRevisionFields": [
"df.019a0000-0000-7000-8000-000000000001"
]
}
}
},
"archivedAt": null,
"createdAt": "2026-01-01T00:00:00Z",
"updatedAt": "2026-01-01T00:00:00Z"
}