Alexlab User Actions Community Edition

The software processes templates and arguments to generate user actions at scale (i.e. prompts for chatbot platforms such as Copilot and Gemini or search queries for TikTok and Youtube).

1. Templates and Arguments: Please see the example_data/ folder for an example structure of templates and arguments. The basic requirement is that each value in the template file may or may not contain a placehoder enclosed in curly brackets, such as "{TOPIC}", and that the arguments file contains one or more arguments, each relevant to one or more different countries, for each of the placeholders mentioned in the templates.

2. Template Interpolation: The software interpolates the templates with the provided arguments. This involves replacing the placeholders in the templates with the corresponding argument values.

3. Prompts and Search Queries Production: The interpolated templates are used to generate user actions which:

   • (translation) may or may not be translated, based on the target countries and languages
   • (stopword removal) may have their stopwords removed through the YAKE algorithm, in case they are to be a search query.

These actions are then saved to a CSV file.

Country implies language, if you will

By default, there is a strong assumption that for each target country actions should be generated only in English (if it is the original language of the template) and in the language whose code corresponds to the country code (e.g. es for Spanish and Spain). You change change this behaviour by using the --ecl parameter.

Translation

By default, to perform translations the module will use the HuggingFace Space aiforensics/opus-mt-translation-ce which hosts a Gradio service to run the open-source machine translation Opus-MT models, developed by the Language Technology Research Group at the University of Helsinki, through the EasyNMT library.

To make translations quicker, you can either:

• make your own private clone of our HF Space, assign additional resources to it, rather than the default CPU, and then use the --hf-space and --hf-token to access it.
• run with Docker (see below).

How to install

git clone https://github.com/aiforensics/alexlab-user-actions-ce.git

cd alexlab-user-actions-ce

pip install -r requirements.x

pip install .

How to run

python -m alexlab_user_actions --input-templates 'example_data/templates.csv' --input-arguments 'example_data/arguments.csv' --output './output.csv'

Run translations with Docker (optional)

Please ensure you have nvidia-docker2 installed (see here). Then run:

docker run -it -p 7860:7860 --platform=linux/amd64 --gpus all \
	registry.hf.space/aiforensics-opus-mt-translation-ce:gpu-3e6cebd python app.py

Use localhost:7860 as the URL for --hf-space.

Options

--input-templates INPUT_TEMPLATES (required)

Path to the input templates CSV file.

--input-arguments INPUT_ARGUMENTS (required)

Path to the input arguments CSV file.

--output OUTPUT (required)

Path to the output CSV file.

--dry-run | --no-dry-run

Don't generate the actions, only report the changes that would be made by running the input files

--translate | --no-translate

Use the translation service (default: true)

--hf-space HF_SPACE

HuggingFace Space to use for translations (optional)

--hf-token HF_TOKEN

HuggingFace Token to use for translations (optional)

--ecl ECL [ECL ...]

List of target extra country-language pairs for the user actions in (see the "Country implies language section").

E.g., to generate for example French and Dutch translations for Belgium), use --ecl be:fr be:nl.

Funding

This contribution from AI Forensics is funded by a project grant from NGI Search.