Syphus: Automatic Instruction-Response Generation Pipeline

Paper | Documents | PyPI | Source Code

Installation

pip install --upgrade syphus

CLI Usage

You can easily use the Syphus CLI for operations.

First, choose an empty folder for initialization:

syphus init <folder>

The structure of the folder will be:

<folder>/
    config/
        gpt_info.yaml
        prompt.yaml
    resources/
        media_infos.json

Configuration

You can edit gpt_info.yaml and prompt.yaml to customize the generation.

Edit Media Data

Then, change the resources/media_infos.json to include the media data you want to generate.

You can use two different styles of resources/media_infos.json:

{
    "id_1": {
        "data_1": "data_1_1",
        "data_2": {
            "data_2_1": "data_2_1_1",
            "data_2_2": "data_2_2_1",
        }
    },
    "id_2": [
        "data_2_1",
        "data_2_2",
    ],
    "id_3": "data_3",
    ...
}

or

[
    {
        "id": "id_1",
        "data_1": "data_1",
        "data_2": {
            "data_2_1": "data_2_1",
            "data_2_2": "data_2_2",
        }
    },
    {
        "id": "id_2",
        "data_1": "data_1",
        "data_2": "data_2",
    },
    ...
]

Optionally, you can change the resources/media_infos.json to resources/media_infos.jsonl, but the format should like the second one of resources/media_infos.json:

{"id": "id_1", "data_1": "data_1", "data_2": "data_2"}
{"id": "id_2", "data_1": "data_1", "data_2": "data_2"}

Generate

After editing the resources/media_infos.json, you can generate the instruction-response pairs:

syphus query <folder>

or

cd <folder>
syphus query

You can use syphus query --help to see more options.