Start Job

POST

training

grpo

import PiClient from 'withpi';

const client = new PiClient({
  apiKey: process.env['WITHPI_API_KEY'], // This is the default and can be omitted
});

async function main() {
  const response = await client.training.grpo.startJob({
    base_rl_model: 'LLAMA_3.2_3B',
    examples: [{ llm_input: 'Tell me something different' }],
    learning_rate: 0.000005,
    lora_config: {},
    num_train_epochs: 10,
    scoring_spec: {
      description: "Write a children's story communicating a simple life lesson.",
      dimensions: [
        {
          description: 'dimension1 description',
          label: 'dimension1',
          sub_dimensions: [
            { description: 'subdimension1 description', label: 'subdimension1', scoring_type: 'PI_SCORER' },
          ],
        },
      ],
      name: 'Sample Scoring Spec',
    },
    system_prompt: 'An optional system prompt.',
  });

  console.log(response.job_id);
}

main();

{
  "detailed_status": [
    "Downloading model",
    "Tuning prompt"
  ],
  "job_id": "1234abcd",
  "state": "RUNNING",
  "trained_models": [
    {
      "epoch": 123,
      "eval_loss": 123,
      "pi_score": 0,
      "serving_id": 123,
      "serving_state": "UNLOADED",
      "step": 123
    }
  ]
}

Authorizations

x-api-key

string

header

required

Body

application/json

base_rl_model

enum<string>

required

The base model to start the RL tunning process

Available options:

LLAMA_3.2_3B,

LLAMA_3.1_8B

examples

object[]

required

Examples to use in the RL tuning process

An example for RL training

learning_rate

number

required

GRPO learning rate

Example:

0.000005

lora_config

object

required

The LoRA configuration.

num_train_epochs

integer

required

GRPO number of train epochs

Example:

10

scoring_spec

object

required

The scoring spec to use in the GRPO tuning process

scoring_spec.description

string

required

The application description

Example:

"Write a children's story communicating a simple life lesson."

scoring_spec.dimensions

object[]

required

The dimensions of the scoring spec

scoring_spec.dimensions.description

string

required

The description of the dimension

Example:

"Relevance of the response"

scoring_spec.dimensions.label

string

required

The label of the dimension

Example:

"Relevance"

scoring_spec.dimensions.sub_dimensions

object[]

required

The sub dimensions of the dimension

scoring_spec.dimensions.sub_dimensions.description

string

required

The description of the dimension

Example:

"Is the response relevant to the prompt?"

scoring_spec.dimensions.sub_dimensions.label

string

required

The label of the dimension

Example:

"Relevance to Prompt"

scoring_spec.dimensions.sub_dimensions.scoring_type

enum<string>

required

The type of scoring performed for this dimension

Available options:

PI_SCORER,

PYTHON_CODE,

CUSTOM_MODEL_SCORER

scoring_spec.dimensions.sub_dimensions.custom_model_id

string | null

The ID of the custom model to use for scoring. Only relevant for scoring_type of CUSTOM_MODEL_SCORER

Example:

"your-model-id"

scoring_spec.dimensions.sub_dimensions.parameters

number[] | null

The learned parameters for the scoring method. This represents piecewise linear interpolation between [0, 1].

Example:

[
  0.14285714285714285,
  0.2857142857142857,
  0.42857142857142855,
  0.5714285714285714,
  0.7142857142857143,
  0.8571428571428571
]

scoring_spec.dimensions.sub_dimensions.python_code

string | null

The PYTHON code associated the PYTHON_CODE DimensionScoringType.

Example:

"\ndef score(response_text: str, input_text: str, kwargs: dict) -> dict:\n word_count = len(response_text.split())\n if word_count > 10:\n return {\"score\": 0.2, \"explanation\": \"Response has more than 10 words\"}\n elif word_count > 5:\n return{\"score\": 0.6, \"explanation\": \"Response has more than 5 words\"}\n else:\n return {\"score\": 1, \"explanation\": \"Response has 5 or fewer words\"}\n"

scoring_spec.dimensions.sub_dimensions.weight

number | null

The weight of the subdimension. The sum of subdimension weights will be normalized to one internally. A higher weight counts for more when aggregating this subdimension into the parent dimension.

Example:

1

Example:

{
  "description": "subdimension1 description",
  "label": "subdimension1",
  "scoring_type": "PI_SCORER"
}

scoring_spec.dimensions.parameters

number[] | null

The learned parameters for the scoring method. This represents piecewise linear interpolation between [0, 1].

Example:

[
  0.14285714285714285,
  0.2857142857142857,
  0.42857142857142855,
  0.5714285714285714,
  0.7142857142857143,
  0.8571428571428571
]

scoring_spec.dimensions.weight

number | null

The weight of the dimension The sum of dimension weights will be normalized to one internally. A higher weight counts for more when aggregating this dimension is aggregated into the final score.

Example:

1

Example:

[
  {
    "description": "dimension1 description",
    "label": "dimension1",
    "sub_dimensions": [
      {
        "description": "subdimension1 description",
        "label": "subdimension1",
        "scoring_type": "PI_SCORER"
      }
    ]
  }
]

scoring_spec.name

string

required

The name of the scoring spec

Example:

"Sample Scoring Spec"

system_prompt

string | null

required

A custom system prompt to use during the RL tuning process

Example:

"An optional system prompt."

Response

200

application/json

Successful Response

RlGrpoStatus is the status of a RL PPO job.

detailed_status

string[]

required

Detailed status of the job

Example:

["Downloading model", "Tuning prompt"]

job_id

string

required

The job id

Example:

"1234abcd"

state

enum<string>

required

Current state of the job

Available options:

QUEUED,

RUNNING,

DONE,

ERROR,

CANCELLED

trained_models

object[] | null

A list of trained models selected based on the PI score.

List Retrieve

import PiClient from 'withpi';

const client = new PiClient({
  apiKey: process.env['WITHPI_API_KEY'], // This is the default and can be omitted
});

async function main() {
  const response = await client.training.grpo.startJob({
    base_rl_model: 'LLAMA_3.2_3B',
    examples: [{ llm_input: 'Tell me something different' }],
    learning_rate: 0.000005,
    lora_config: {},
    num_train_epochs: 10,
    scoring_spec: {
      description: "Write a children's story communicating a simple life lesson.",
      dimensions: [
        {
          description: 'dimension1 description',
          label: 'dimension1',
          sub_dimensions: [
            { description: 'subdimension1 description', label: 'subdimension1', scoring_type: 'PI_SCORER' },
          ],
        },
      ],
      name: 'Sample Scoring Spec',
    },
    system_prompt: 'An optional system prompt.',
  });

  console.log(response.job_id);
}

main();

{
  "detailed_status": [
    "Downloading model",
    "Tuning prompt"
  ],
  "job_id": "1234abcd",
  "state": "RUNNING",
  "trained_models": [
    {
      "epoch": 123,
      "eval_loss": 123,
      "pi_score": 0,
      "serving_id": 123,
      "serving_state": "UNLOADED",
      "step": 123
    }
  ]
}

Get Started

Client Libraries

API Reference

Authorizations

Body

Response