The gpt-subb from skyatura

Retry failed chunks before throwing

Use completion api over chat api

Chat API is not the best fit for translating only tasks, the tool should use Completion API which has more fine grained options to consistent results.

There is also a new candidate API for replacing this, which is Edit API. This is still Beta but we may end up using it, since this is the exactly use case of it.

"Continue" mode

There must be an option to start from where a previous run stopped.

This is a follow up feature of #8

sometimes getting prompt failed.

Using Arch Linux, node 16, command gpt-subb -m gpt-3.5-turbo -l tr -f SRT tesst/[Judas]\ High\ School\ DxD\ S1\ -\ 01.srt

  prompt: 'Translate the following text into tr but keep the 6 digit codes between < > intact:\n' +
    '\n' +
    '<YRD613>\n' +
    '<font face="ITC Stone Sans Std Medium" size="52"><b>A-A devil?</b></font>\n' +
    '<000000>\n' +
    '<XJK848>\n' +
    '<font face="ITC Stone Sans Std Medium" size="52"><b>And your master.</b></font>\n' +
    '<000000>\n' +
    '<JIQ772>\n' +
    '<font face="ITC Stone Sans Std Medium" size="52"><b>Nice to have you, Hyoudou Issei-kun.</b></font>\n' +
    '<000000>\n' +
    '<IFM795>\n' +
    '<font face="Boopee" size="50" color="#d0d52a"><font color="#fdfdfd"><font color="#f91600">High School DxD</font></font></font>\n' +
    '<000000>\n' +
    '<EQX360>\n' +
    '<font face="Boopee" size="50" color="#d0d52a"><font color="#ff5169"><font color="#f91600">High School DxD</font></font></font>\n' +
    '<000000>\n' +
    '<ZBX292>\n' +
    '<font face="Boopee" size="50" color="#d0d52a"><font color="#ff5064">High School DxD</font></font>\n' +
    '<000000>\n' +
    '<IBN377>\n' +
    '<font face="Boopee" size="50" color="#d0d52a"><font color="#ff5300">High School DxD</font></font>\n' +
    '<000000>\n' +
    '<LUI665>\n' +
    '<font face="Boopee" size="50" color="#d0d52a"><font color="#f19498">High School DxD</font></font>\n' +
    '<000000>\n' +
    '<EAA572>\n' +
    '<font face="Boopee" size="50" color="#d0d52a"><font color="#dcd0f8">High School DxD</font></font>\n' +
    '<000000>\n' +
    '<XGY981>\n' +
    '<font face="Aurulent Sans" size="50">The lines come together with a distant voice calling in the skies,</font>\n' +
    '<000000>\n' +
    '<ROG759>\n' +
    '<font face="Baar Sophia" size="50">{\\an8}</font><font face="ITC Stone Sans Std Medium" size="52"><b><font color="#78595b"><font color="#ffffff">ma</font></font></b></font>\n' +
    '<000000>\n' +
    '<OYX736>\n' +
    '<font face="Baar Sophia" size="50">{\\an8}</font><font face="ITC Stone Sans Std Medium" size="52"><b><font color="#78595b"><font color="#ffffff">ji</font></font></b></font>\n' +
    '<000000>\n' +
    '<FLZ693>\n' +
    '<font face="Baar Sophia" size="50">{\\an8}</font><font face="ITC Stone Sans Std Medium" size="52"><b><font color="#78595b"><font color="#ffffff">wa</font></font></b></font>\n' +
    '<000000>\n' +
    '<OUH089>\n' +
    '<font face="Baar Sophia" size="50">{\\an8}</font><font face="ITC Stone Sans Std Medium" size="52"><b><font color="#78595b"><font color="#ffffff">ri</font></font></b></font>\n' +
    '<000000>\n' +
    '<OJI981>\n' +
    '<font face="Baar Sophia" size="50">{\\an8}</font><font face="ITC Stone Sans Std Medium" size="52"><b><font color="#78595b"><font color="#ffffff">a</font></font></b></font>',
  error: Error: Request failed with status code 502
      at createError (/home/tbb/.nvm/versions/node/v16.20.0/lib/node_modules/gpt-subb/node_modules/openai/node_modules/axios/lib/core/createError.js:16:15)
      at settle (/home/tbb/.nvm/versions/node/v16.20.0/lib/node_modules/gpt-subb/node_modules/openai/node_modules/axios/lib/core/settle.js:17:12)
      at IncomingMessage.handleStreamEnd (/home/tbb/.nvm/versions/node/v16.20.0/lib/node_modules/gpt-subb/node_modules/openai/node_modules/axios/lib/adapters/http.js:322:11)
      at IncomingMessage.emit (node:events:525:35)
      at endReadableNT (node:internal/streams/readable:1358:12)
      at processTicksAndRejections (node:internal/process/task_queues:83:21) {
    config: {
      transitional: [Object],
      adapter: [Function: httpAdapter],
      transformRequest: [Array],
      transformResponse: [Array],
      timeout: 0,
      xsrfCookieName: 'XSRF-TOKEN',
      xsrfHeaderName: 'X-XSRF-TOKEN',
      maxContentLength: -1,
      maxBodyLength: -1,
      validateStatus: [Function: validateStatus],
      headers: [Object],
      method: 'post',
      data: '{"model":"gpt-3.5-turbo","messages":[{"content":"Translate the following text into tr but keep the 6 digit codes between < > intact:\\n\\n<YRD613>\\n<font face=\\"ITC Stone Sans Std Medium\\" size=\\"52\\"><b>A-A devil?</b></font>\\n<000000>\\n<XJK848>\\n<font face=\\"ITC Stone Sans Std Medium\\" size=\\"52\\"><b>And your master.</b></font>\\n<000000>\\n<JIQ772>\\n<font face=\\"ITC Stone Sans Std Medium\\" size=\\"52\\"><b>Nice to have you, Hyoudou Issei-kun.</b></font>\\n<000000>\\n<IFM795>\\n<font face=\\"Boopee\\" size=\\"50\\" color=\\"#d0d52a\\"><font color=\\"#fdfdfd\\"><font color=\\"#f91600\\">High School DxD</font></font></font>\\n<000000>\\n<EQX360>\\n<font face=\\"Boopee\\" size=\\"50\\" color=\\"#d0d52a\\"><font color=\\"#ff5169\\"><font color=\\"#f91600\\">High School DxD</font></font></font>\\n<000000>\\n<ZBX292>\\n<font face=\\"Boopee\\" size=\\"50\\" color=\\"#d0d52a\\"><font color=\\"#ff5064\\">High School DxD</font></font>\\n<000000>\\n<IBN377>\\n<font face=\\"Boopee\\" size=\\"50\\" color=\\"#d0d52a\\"><font color=\\"#ff5300\\">High School DxD</font></font>\\n<000000>\\n<LUI665>\\n<font face=\\"Boopee\\" size=\\"50\\" color=\\"#d0d52a\\"><font color=\\"#f19498\\">High School DxD</font></font>\\n<000000>\\n<EAA572>\\n<font face=\\"Boopee\\" size=\\"50\\" color=\\"#d0d52a\\"><font color=\\"#dcd0f8\\">High School DxD</font></font>\\n<000000>\\n<XGY981>\\n<font face=\\"Aurulent Sans\\" size=\\"50\\">The lines come together with a distant voice calling in the skies,</font>\\n<000000>\\n<ROG759>\\n<font face=\\"Baar Sophia\\" size=\\"50\\">{\\\\an8}</font><font face=\\"ITC Stone Sans Std Medium\\" size=\\"52\\"><b><font color=\\"#78595b\\"><font color=\\"#ffffff\\">ma</font></font></b></font>\\n<000000>\\n<OYX736>\\n<font face=\\"Baar Sophia\\" size=\\"50\\">{\\\\an8}</font><font face=\\"ITC Stone Sans Std Medium\\" size=\\"52\\"><b><font color=\\"#78595b\\"><font color=\\"#ffffff\\">ji</font></font></b></font>\\n<000000>\\n<FLZ693>\\n<font face=\\"Baar Sophia\\" size=\\"50\\">{\\\\an8}</font><font face=\\"ITC Stone Sans Std Medium\\" size=\\"52\\"><b><font color=\\"#78595b\\"><font color=\\"#ffffff\\">wa</font></font></b></font>\\n<000000>\\n<OUH089>\\n<font face=\\"Baar Sophia\\" size=\\"50\\">{\\\\an8}</font><font face=\\"ITC Stone Sans Std Medium\\" size=\\"52\\"><b><font color=\\"#78595b\\"><font color=\\"#ffffff\\">ri</font></font></b></font>\\n<000000>\\n<OJI981>\\n<font face=\\"Baar Sophia\\" size=\\"50\\">{\\\\an8}</font><font face=\\"ITC Stone Sans Std Medium\\" size=\\"52\\"><b><font color=\\"#78595b\\"><font color=\\"#ffffff\\">a</font></font></b></font>","role":"user"}],"temperature":0.4}',
      url: 'https://api.openai.com/v1/chat/completions'
    },
    request: ClientRequest {
      _events: [Object: null prototype],
      _eventsCount: 7,
      _maxListeners: undefined,
      outputData: [],
      outputSize: 0,
      writable: true,
      destroyed: false,
      _last: true,
      chunkedEncoding: false,
      shouldKeepAlive: false,
      maxRequestsOnConnectionReached: false,
      _defaultKeepAlive: true,
      useChunkedEncodingByDefault: true,
      sendDate: false,
      _removedConnection: false,
      _removedContLen: false,
      _removedTE: false,
      strictContentLength: false,
      _contentLength: 2478,
      _hasBody: true,
      _trailer: '',
      finished: true,
      _headerSent: true,
      _closed: false,
      socket: [TLSSocket],
      _header: 'POST /v1/chat/completions HTTP/1.1\r\n' +
        'Accept: application/json, text/plain, */*\r\n' +
        'Content-Type: application/json\r\n' +
        'User-Agent: OpenAI/NodeJS/3.2.1\r\n' +
        'Authorization: Bearer (they were key )n' +
        'Content-Length: 2478\r\n' +
        'Host: api.openai.com\r\n' +
        'Connection: close\r\n' +
        '\r\n',
      _keepAliveTimeout: 0,
      _onPendingData: [Function: nop],
      agent: [Agent],
      socketPath: undefined,
      method: 'POST',
      maxHeaderSize: undefined,
      insecureHTTPParser: undefined,
      path: '/v1/chat/completions',
      _ended: true,
      res: [IncomingMessage],
      aborted: false,
      timeoutCb: null,
      upgradeOrConnect: false,
      parser: null,
      maxHeadersCount: null,
      reusedSocket: false,
      host: 'api.openai.com',
      protocol: 'https:',
      _redirectable: [Writable],
      [Symbol(kCapture)]: false,
      [Symbol(kBytesWritten)]: 0,
      [Symbol(kEndCalled)]: true,
      [Symbol(kNeedDrain)]: false,
      [Symbol(corked)]: 0,
      [Symbol(kOutHeaders)]: [Object: null prototype],
      [Symbol(kUniqueHeaders)]: null
    },
    response: {
      status: 502,
      statusText: 'Bad Gateway',
      headers: [Object],
      config: [Object],
      request: [ClientRequest],
      data: '<html>\r\n' +
        '<head><title>502 Bad Gateway</title></head>\r\n' +
        '<body>\r\n' +
        '<center><h1>502 Bad Gateway</h1></center>\r\n' +
        '<hr><center>cloudflare</center>\r\n' +
        '</body>\r\n' +
        '</html>\r\n'
    },
    isAxiosError: true,
    toJSON: [Function: toJSON]
  }
}
Translating queue 25 of 35... failed! (skipping)

[1.1.x] Refactor the entire tool

Since I first created this project, my understanding on OpenAI's API and LLMs got significantly improved. That being said, I think this deserve an update for implementing things I didn't knew back then, alongside with better error handling.

Goals for the new version

Multiline support is missing

When subtitles are on multiples lines, only the first one is written back in the subtitle file.

Translations queues:

  '<VPE875>\n- I like you.\n- I like you.',
  '<STX511>\nSome unexpected costs have come up.',
  '<YVT450>\n- What do you need?\n- $50,000.',
  '<GHN797>\nYou bet the money that my\nfather gave us to pay the lease?',
  "<VMJ674>\nRANJIT: It's gone, man. It's all gone.",
  "<DVU397>\nMALIKA: Why did Jack Hauss\nvote against the women's center",
  '<BPI409>\nwhen he told me, to my face,\nthat he would support it?',
[...]

srt file:

1
00:00:00,658 --> 00:00:02,704
- I like you.

2
00:00:08,927 --> 00:00:11,582
Some unexpected costs have come up.

3
00:00:11,626 --> 00:00:14,150
- What do you need?

4
00:00:14,413 --> 00:00:17,155
You bet the money that my

5
00:00:17,199 --> 00:00:18,766
RANJIT: It's gone, man. It's all gone.

6
00:00:18,809 --> 00:00:20,724
MALIKA: Why did Jack Hauss

7
00:00:20,768 --> 00:00:22,595
when he told me, to my face,

[...]

Budget limit

TL;DR: OpenAI is draining out my money while I attempt designing better prompts, and I need to pause for a while.

I finally took some time to improve this project, and I have many things in mind for it, as indicated by the issues I've tracked earlier today.

However, even though this is an open source project and I don't pretend to charge for it, developing with OpenAI's API can become really expensive in just a few minutes. Specially because subtitles are large files which consume many tokens for each run, and designing consistent prompts require a lot of trials.

This being said, since I already reached my monthly budget, I will need to pause the development until next billing cycle.

Better default prompt

A well designed prompt helps to get consistent result over many iterations.

The prompt must guarantee:

The cue reference must not be modified
The whole cue will be used in the translation
The cue line count must match the original one whenever possible
Surrounding cues could be used as context for better translation
The prompt could allow users to add additional instructions, like
- Avoid bad words
- Explain technical terms
- Explain abbreviations

Bonus features

I will also experiment on things like adding cultural context on jokes and references that may be lost in translation

Better chunk management

TypeError

I can't seem to get this working properly, neither in Windows, nor with Colab.
The output is the same in both, any pointers towards the right direction are greatly appreciated.

Converting translated cues to objects...
TypeError: undefined is not iterable (cannot read property Symbol(Symbol.iterator))
    at /usr/local/lib/node_modules/gpt-subb/dist/index.js:167:33
    at Array.map (<anonymous>)
    at translatedToCues (/usr/local/lib/node_modules/gpt-subb/dist/index.js:166:29)
    at /usr/local/lib/node_modules/gpt-subb/dist/index.js:202:28
    at Generator.next (<anonymous>)
    at fulfilled (/usr/local/lib/node_modules/gpt-subb/dist/index.js:29:58)
    at process.processTicksAndRejections (node:internal/process/task_queues:95:5)

Only the first line of multiline subtitles are translated.

Input:

12
00:03:06,733 --> 00:03:11,832

¿Está a 2000 metros de altitud?
¡No, doctor!

13
00:03:11,852 --> 00:03:15,871

¡Quizá 1950!
¿Se puede llegar allí con Jeep?

Output:

12
00:03:06,733 --> 00:03:11,832

2000 metre yükseklikte mi?

13
00:03:11,852 --> 00:03:15,871

Belki 1950!

skyatura / gpt-subb Goto Github PK

gpt-subb's Introduction

GPT Subb

First things first

Installation

Usage

Prerequisites

Command

Options

Environment variables

Examples

gpt-subb's People

Contributors

Stargazers

Watchers

Forkers

gpt-subb's Issues

Goals for the new version

Bonus features

Recommend Projects

Recommend Topics

Recommend Org