Template: synthetic

synthetic is a datafact template for creating datasets by generating synthetic using LLM. This template make use of another library mkb.

In order to proceed with this template, you need to setup mkb according to this guide

Create a new project using this template

datafact new datafact-tutorial/synthetic -t synthetic

Click to show output

Now go to the project folder datafact-tutorial/synthetic.

cd datafact-tutorial/synthetic

ls

Hide Output

README.md         data.py           datafact.json     fn.py             project.py        type.py

What to modify

You should modify the following files:

• data.py
• type.py
• fn.py
• README.md

data.py

Implement a function called create_data_dict which produce the content of your dataset.

import sys
from dataset_sh.constants import DEFAULT_COLLECTION_NAME
from fn import generate_greeting_message


def create_main_collection():
    languages = [
        "English",
        "Spanish",
        "Mandarin Chinese",
        "Hindi",
        "Arabic",
        "Portuguese",
        "Bengali",
        "Russian",
        "Japanese",
        "Punjabi",
        "German",
        "Javanese",
        "Korean",
        "French",
        "Telugu",
        "Vietnamese",
        "Turkish",
        "Italian",
        "Tamil",
        "Urdu"
    ]

    batch = generate_greeting_message.batch('generate_greeting_message_batch_0')

    if batch.can_start_batch():
        for lang in languages:
            batch.add(lang)

        batch.start_batch()
        print('Batch started, please wait until it was processed.')
        sys.exit(1)
    else:
        print('batch already started.')
        status = batch.sync_remote()
        print(status)
        if status['status'] == 'finished':
            items = []
            for input_args, output in batch.iter_outputs():
                items.append({
                    'language': input_args['args'][0],
                    "value": output.value
                })
            return items
        else:
            print('batch is not finished, exit dataset building process now.')
            sys.exit(1)


def create_data_dict():
    return {
        DEFAULT_COLLECTION_NAME: create_main_collection()
    }

fn.py

Provide your mkb function definition.

import sys
import os
import click
from openai import OpenAI
from mkb import def_fn

mkb_openai_api_key = os.environ.get('MKB_OPENAI_API_KEY', None)

if mkb_openai_api_key is None:
    open_ai_key_file = os.path.expanduser('~/.openai.key')
    if os.path.exists(open_ai_key_file):
        with open(open_ai_key_file) as f:
            mkb_openai_api_key = f.read().strip()

if mkb_openai_api_key is None or mkb_openai_api_key == '':
    click.secho('Please set MKB_OPENAI_API_KEY environment variable or ~/.openai.key file.', fg='red')
    sys.exit(1)

from mkb.impl.openai import GptApiOptions

openai_client = OpenAI(api_key=mkb_openai_api_key)

my_gpt_opts = GptApiOptions(temperature=1.0)


@def_fn.openai(client=openai_client, opts=my_gpt_opts)
@def_fn.with_instruction("Given a language, generate a greeting message in that language.")
@def_fn.with_example(
    input='English',
    output='Hello'
)
@def_fn.with_example(
    input='Chinese',
    output='你好'
)
def generate_greeting_message(lang):
    # any code you wrote here will be ignored.
    pass


if __name__ == "__main__":
    names = generate_greeting_message('Spanish')
    print(names.value)

type.py

Provide your type annotation in the data_types dictionary in this file.

from easytype import TypeBuilder
from dataset_sh.constants import DEFAULT_COLLECTION_NAME

GreetingMessage = TypeBuilder.create(
    'GreetingMessage',
    language=str,
    value=str
)

data_types = {
    DEFAULT_COLLECTION_NAME: GreetingMessage
}

README.md

Provide the readme document of your dataset in this file.

# datafact template: synthetic

This dataset is built by the datafact template: `synthetic`,

# Create and publish dataset

This folder contains a project created by datafact (DATAset FACTory),

You can use this project to build and publish your dataset to [dataset.sh](https://dataset.sh)

## How to modify this project


You may want to modify the following files as follows:

* data.py
    Implement a function called `create_data_dict` which produce the content of your dataset.

* type.py
    Provide your type annotation in the `data_types` dictionary in this file.

* fn.py
    Provide your mkb function definition.

* README.md
    Provide the readme document of your dataset in this file.

## Build Dataset

To build your dataset, use command:

```shell
python project.py build
```

## Publish Dataset

To publish your dataset, use command:

```shell
python project.py publish
```

You can also using the experimental auto-doc feature, to generated a documentation based on the type annotation stored in type.py:

datafact experimental auto-doc

Build and publish

To build your dataset

python project.py build

To preview your build before publish

python project.py preview show

To publish your build

python project.py publish

Example: science/acl-datasets

More Examples