slang_gpt 0.1.0

Use GPT to automatically translate at compile time.

slang_gpt

Use GPT to automatically translate your app at compile time.

This is library is intended to be used with slang.

Currently, only the OpenAI API is supported.

Motivation

Google Translate and other translation services are great, but they are not perfect.

One of the biggest issues is that they are not context aware. For example, the word "bank" can be translated to "Bank" or "Ufer" in German depending on the context.

With GPT and some prompt engineering, we can get context aware translations.

Getting Started

# pubspec.yaml
dependencies:
  slang: <version>

dev_dependencies:
  slang_gpt: <version>

Add the gpt entry to build.yaml or slang.yaml.

# existing config
base_locale: fr
fallback_strategy: base_locale
input_directory: lib/i18n
input_file_pattern: .i18n.json
output_directory: lib/i18n

# add this
gpt:
  model: gpt-3.5-turbo
  description: |
    "River Adventure" is a game where you need to cross a river by jumping on stones.
    The game is over when you either fall into the water or reach the other side.

Let's run this:

dart run slang_gpt --target=fr --api-key=<api-key>

Configuration

Key	Type	Usage	Required	Default
model	String	Model name	YES
max_input_length	int	Max input characters per request	NO	(inferred by model)
description	String	App description	YES

Command line arguments

Argument	Description	Required	Default
--target=	Target language	NO	(all existing locales)
--api-key=	API key	YES
-f / --full	Skip partial translation	NO	(partial translation)
-v / --verbose	Write chat to file	NO	(not verbose)
--outdir=	Output directory	NO*	(using config)

Models

Model name	Provider	Context length	Cost per 1k input token	Cost per input word (English)
gpt-3.5-turbo	Open AI	4000	$0.0015	$0.000001125
gpt-3.5-turbo-16k	Open AI	16000	$0.003	$0.00000225

GPT context length

Each model has a different context length. Try to avoid exceeding it as the model starts to "forget".

Luckily, slang_gpt supports splitting the input into multiple requests.

The max_input_length is optional and defaults to some heuristic.

If you work with less common languages and the model starts to forget, try to reduce the max_input_length.

Alternatively, you can also use a model with a larger context length.

Partial translation

By default, slang_gpt will only translate missing keys to reduce costs.

You may add the --full flag to translate all keys.

Target language

By default, slang_gpt will translate to all existing locales.

You may add the --target flag to translate to a specific locale. This may be useful if you want to translate to a new locale.

Additionally, you may also use predefined language sets (keep in mind that English must be the base locale):

By GDP (Gross Domestic Product):

Flag	Languages
--target=gdp-3	["zh-Hans", "es", "ja"]
--target=gdp-5	["zh-Hans", "es", "ja", "de", "fr"]
--target=gdp-10	["zh-Hans", "es", "ja", "de", "fr", "pt", "ar", "it", "ru", "ko"]

By region and population:

Flag	Languages
--target=eu-3	["de", "fr", "it"]
--target=eu-5	["de", "fr", "it", "es", "pl"]
--target=eu-10	["de", "fr", "it", "es", "pl", "ro", "nl", "cs", "el", "sv"]