OpenAI #FtJob: 生成

// GraalJS standard mode Script Script (engine type: 3)
// cf. 'engine type: 2': "GraalJS Nashorn compatible mode" (renamed from "GraalJS" at 20230526)


//////// START "main()" /////////////////////////////////////////////////////////////////

main();
function main(){ 

////// == Config Retrieving / 工程コンフィグの参照 ==
const strAuthzSetting     = configs.get( "AuthzConfU1" );             /// REQUIRED
  engine.log( " AutomatedTask Config: Authz Setting: " + strAuthzSetting );
const strOrgId            = configs.get( "StrConfU2" );               // NotRequired
  engine.log( " AutomatedTask Config: OpenAI-Organization: " + strOrgId );

const strFileIdTraining   = configs.get( "StrConfA1" );               /// REQUIRED
const strFileIdValidation = configs.get( "StrConfA2" );               // NotRequired
const strBaseModel        = configs.get( "StrConfA3" );               /// REQUIRED (20230822)
const strSuffix           = configs.get( "StrConfA4" );               // NotRequired

const strPocketJobId      = configs.getObject( "SelectConfB1" );      /// REQUIRED



////// == Data Retrieving / ワークフローデータの参照 ==
// (Nothing. Retrieved via Expression Language in Config Retrieving)



////// == Calculating / 演算 ==

//// OpenAI API > Documentation > API REFERENCE > Fine-tunes > Create fine-tuning job
//// https://platform.openai.com/docs/api-reference/fine-tuning/create

/// prepare json
let strJson = {};
    strJson.training_file = strFileIdTraining;
    if ( strFileIdValidation !== "" ){
      strJson.validation_file = strFileIdValidation;
    }
    strJson.model = strBaseModel;   // eg. "gpt-3.5-turbo" (asof 20230822)
    if ( strSuffix !== "" ){
      strJson.suffix = strSuffix;
    }

/// prepare request1
let request1Uri = "https://api.openai.com/v1/fine_tuning/jobs"; // updated 20230822
let request1 = httpClient.begin(); // HttpRequestWrapper
    request1 = request1.authSetting( strAuthzSetting ); // with "Authorization: Bearer XX"
    if ( strOrgId !== "" ){
      request1 = request1.header( "OpenAI-Organization", strOrgId );
    }
    request1 = request1.body( JSON.stringify( strJson ), "application/json" );

/// try request1
const response1     = request1.post( request1Uri ); // HttpResponseWrapper
engine.log( " AutomatedTask ApiRequest1 Start: " + request1Uri );
const response1Code = response1.getStatusCode() + ""; // JavaNum to string
const response1Body = response1.getResponseAsString();
engine.log( " AutomatedTask ApiResponse1 Status: " + response1Code );
if( response1Code !== "200"){
  throw new Error( "\n AutomatedTask UnexpectedResponseError: " +
                    response1Code + "\n" + response1Body + "\n" );
}

/* engine.log( response1Body ); // debug
{
  "object":"fine_tuning.job",
  "id":"ftjob-EcRIIXZMnZzbEqyrFvJ2vYLZ",
  "model":"gpt-3.5-turbo-0613",
  "created_at":1692772480,
  "finished_at":null,
  "fine_tuned_model":null,
  "organization_id":"org-xxxxxyyyyyzzzzzXXXXXYYYY",
  "result_files":[],
  "status":"created",
  "validation_file":null,
  "training_file":"file-UaXSB3M90mpinC3pH3ZGiTiJ",
  "hyperparameters":{"n_epochs":5},
  "trained_tokens":null
}
*/

/// parse response1
const response1Obj = JSON.parse( response1Body );



////// == Data Updating / ワークフローデータへの代入 ==
if( strPocketJobId !== null ){
  engine.setData( strPocketJobId, response1Obj.id );
}


} //////// END "main()" /////////////////////////////////////////////////////////////////


/*
Notes:
- This "Automated Step" will start to generate a Fine-Tuned MODEL on the OpenAI server.
    - It is possible to specify the base MODEL.
    - The confirmation step is placed downstream. (OpenAI #FtJob: Retrieve)
- If you place this "Automated Step" in the Workflow diagram, the request will be automatically sent every time the process token arrives.
    - A request is automatically sent to the OpenAI API server. (REST API)
    - The response from the OpenAI API server is automatically parsed.

APPENDIX
- To activate a Workflow App that includes this Automated Step, "HTTP Authz Setting" is required
    - Obtain a "Secret API Key" in advance.
    - Set the key as the communication token in "Token Fixed Value"
- This automated step does not support the following parameters.
     - `n_epochs`
     - `batch_size`
     - `learning_rate_multiplier`
     - `prompt_loss_weight`
     - `compute_classification_metrics`
     - `classification_n_classes`
     - `classification_positive_class`
     - `classification_betas`
- Check the OpenAI document for the MODEL that can be specified as the Base MODEL for fine tuning.
    - https://platform.openai.com/docs/guides/fine-tuning/what-models-can-be-fine-tuned
    - Asof 202307: Fine-tuning is currently only available for the following base models: `davinci`, `curie`, `babbage`, and `ada`. We are working on safely enabling fine-tuning for GPT-4 and GPT-3.5 Turbo and expect this feature to be available later this year.
    - Asof 202308: Fine-tuning is currently available for the following models:
        - gpt-3.5-turbo-0613 (recommended)
        - babbage-002
        - davinci-002
- When setting SUFFIX, set it within 40 characters.
     - eg: "custom-model-name"
     - MODEL name: "ada:ft-your-org:custom-model-name-2022-02-15-04-21-04"
- The format of the training data has also changed. (2023-08-22)
    - https://platform.openai.com/docs/guides/fine-tuning/preparing-your-dataset
    - 20230822~: `{"messages": [{"role": "system", "content": "●"}, {"role": "user", "content": "●"}, {"role": "assistant", "content": "●"}]}`
    - ~20230822: `{"prompt": "<prompt text>", "completion": "<ideal generated text>"}`
    - error to: `"message": "File file-xxxxxyyyyyzzzzzXXXXXYYYY format is not compatible with model gpt-3.5-turbo-0613"`
    - At least 10 examples are required. For clear improvements on 50 to 100 training examples typically.
    - error to: `"message": "file-xxxxxyyyyyzzzzzXXXXXYYYY has 4 example(s), but must have at least 10 examples"`


Notes-ja:
- この［自動工程］は、OpenAI サーバ上の Fine-Tuned MODEL 生成を開始させます。
    - ベースとなる MODEL を指定できます。
    - ファインチューニング（微調整）の成果は、下流に進捗確認工程を配置し確認します。（OpenAI #FtJob: 進捗確認）
- この［自動工程］をワークフロー図に配置すれば、案件が到達する度にリクエストが自動送信されます。
    - OpenAI API サーバに対してリクエストが自動送出されます。（REST API通信）
    - OpenAI API サーバからのレスポンスが自動保存解析されます。

APPENDIX-ja
- この［アドオン自動工程］を含むワークフローアプリを運用するには［HTTP 認証設定］が必要です。
    - あらかじめ "Secret API Key" を取得しておいてください。
    - "Secret API Key" を通信トークンとしてセットします。［トークン直接指定］
- この自動工程は、以下のパラメータに対応していません。
    - `n_epochs`
    - `batch_size`
    - `learning_rate_multiplier`
    - `prompt_loss_weight`
    - `compute_classification_metrics`
    - `classification_n_classes`
    - `classification_positive_class`
    - `classification_betas`
- ファインチューニングの Base MODEL として指定可能な MODEL については OpenAI ドキュメントを参照ください。
    - https://platform.openai.com/docs/guides/fine-tuning/what-models-can-be-fine-tuned
    - Asof 202307: Fine-tuning is currently only available for the following base models: `davinci`, `curie`, `babbage`, and `ada`. We are working on safely enabling fine-tuning for GPT-4 and GPT-3.5 Turbo and expect this feature to be available later this year.
    - Asof 202308: Fine-tuning is currently available for the following models:
        - gpt-3.5-turbo-0613 (recommended)
        - babbage-002
        - davinci-002
- SUFFIX を設定する場合は40文字以内で設定します。
    - eg: "custom-model-name"
    - MODEL name: "ada:ft-your-org:custom-model-name-2022-02-15-04-21-04"
- トレーニングデータのフォーマットも変更されています。 (2023-08-22)
    - https://platform.openai.com/docs/guides/fine-tuning/preparing-your-dataset
    - 20230822~: `{"messages": [{"role": "system", "content": "●"}, {"role": "user", "content": "●"}, {"role": "assistant", "content": "●"}]}`
    - ~20230822: `{"prompt": "<prompt text>", "completion": "<ideal generated text>"}`
    - error to: `"message": "File file-xxxxxyyyyyzzzzzXXXXXYYYY format is not compatible with model gpt-3.5-turbo-0613"`
    - 少なくとも10個のトレーニングデータが必要です。通常、50～100個のデータで明確な改善が期待されます。
        - error to: `"message": "file-xxxxxyyyyyzzzzzXXXXXYYYY has 4 example(s), but must have at least 10 examples"`
*/