methods.ts

// File generated from our OpenAPI spec by Stainless. See CONTRIBUTING.md for details.

import { APIResource } from '../../core/resource';
import * as GraderModelsAPI from '../graders/grader-models';

export class Methods extends APIResource {}

/**
 * The hyperparameters used for the DPO fine-tuning job.
 */
export interface DpoHyperparameters {
  /**
   * Number of examples in each batch. A larger batch size means that model
   * parameters are updated less frequently, but with lower variance.
   */
  batch_size?: 'auto' | number;

  /**
   * The beta value for the DPO method. A higher beta value will increase the weight
   * of the penalty between the policy and reference model.
   */
  beta?: 'auto' | number;

  /**
   * Scaling factor for the learning rate. A smaller learning rate may be useful to
   * avoid overfitting.
   */
  learning_rate_multiplier?: 'auto' | number;

  /**
   * The number of epochs to train the model for. An epoch refers to one full cycle
   * through the training dataset.
   */
  n_epochs?: 'auto' | number;
}

/**
 * Configuration for the DPO fine-tuning method.
 */
export interface DpoMethod {
  /**
   * The hyperparameters used for the DPO fine-tuning job.
   */
  hyperparameters?: DpoHyperparameters;
}

/**
 * The hyperparameters used for the reinforcement fine-tuning job.
 */
export interface ReinforcementHyperparameters {
  /**
   * Number of examples in each batch. A larger batch size means that model
   * parameters are updated less frequently, but with lower variance.
   */
  batch_size?: 'auto' | number;

  /**
   * Multiplier on amount of compute used for exploring search space during training.
   */
  compute_multiplier?: 'auto' | number;

  /**
   * The number of training steps between evaluation runs.
   */
  eval_interval?: 'auto' | number;

  /**
   * Number of evaluation samples to generate per training step.
   */
  eval_samples?: 'auto' | number;

  /**
   * Scaling factor for the learning rate. A smaller learning rate may be useful to
   * avoid overfitting.
   */
  learning_rate_multiplier?: 'auto' | number;

  /**
   * The number of epochs to train the model for. An epoch refers to one full cycle
   * through the training dataset.
   */
  n_epochs?: 'auto' | number;

  /**
   * Level of reasoning effort.
   */
  reasoning_effort?: 'default' | 'low' | 'medium' | 'high';
}

/**
 * Configuration for the reinforcement fine-tuning method.
 */
export interface ReinforcementMethod {
  /**
   * The grader used for the fine-tuning job.
   */
  grader:
    | GraderModelsAPI.StringCheckGrader
    | GraderModelsAPI.TextSimilarityGrader
    | GraderModelsAPI.PythonGrader
    | GraderModelsAPI.ScoreModelGrader
    | GraderModelsAPI.MultiGrader;

  /**
   * The hyperparameters used for the reinforcement fine-tuning job.
   */
  hyperparameters?: ReinforcementHyperparameters;
}

/**
 * The hyperparameters used for the fine-tuning job.
 */
export interface SupervisedHyperparameters {
  /**
   * Number of examples in each batch. A larger batch size means that model
   * parameters are updated less frequently, but with lower variance.
   */
  batch_size?: 'auto' | number;

  /**
   * Scaling factor for the learning rate. A smaller learning rate may be useful to
   * avoid overfitting.
   */
  learning_rate_multiplier?: 'auto' | number;

  /**
   * The number of epochs to train the model for. An epoch refers to one full cycle
   * through the training dataset.
   */
  n_epochs?: 'auto' | number;
}

/**
 * Configuration for the supervised fine-tuning method.
 */
export interface SupervisedMethod {
  /**
   * The hyperparameters used for the fine-tuning job.
   */
  hyperparameters?: SupervisedHyperparameters;
}

export declare namespace Methods {
  export {
    type DpoHyperparameters as DpoHyperparameters,
    type DpoMethod as DpoMethod,
    type ReinforcementHyperparameters as ReinforcementHyperparameters,
    type ReinforcementMethod as ReinforcementMethod,
    type SupervisedHyperparameters as SupervisedHyperparameters,
    type SupervisedMethod as SupervisedMethod,
  };
}