title	emoji	colorFrom	colorTo	sdk	sdk_version	app_file	pinned
Alpaca-Guarani	🦙 🇵🇾	green	blue	gradio	5.11.0	main.py	false

Curate Alpaca

About

Simple tool to curate alpaca dataset for training LLMs in other language.

In this case we built this tool to curate dataset transalated to Guarani and prepare for finetunning LLMs with this new language.

Key Features

• Use Alpaca based instruction dataset
• Using nllb-200-distilled-600M to translate into the desired language (Guaraní in this case)
• Supabase for manage the dataset
• Gradio to build the UI
• HF Transformers pipeline to translate the dataset
• Multiple translation backends supported:
  • OpenAI GPT-4 (default) for high-quality translations
  • NLLB (Neural Language Learning Benchmark) model for offline/local processing
  • Google Cloud Translation for reliable machine translation

How to use

1. Clone the repo

git clone this repo

2. Install the requirements

pip install -r requirements.txt

3. Set env vars.

export SUPABASE_URL=your_supabase_url
export SUPABASE_KEY=your_supabase_key
export OPENAI_API_KEY=your_openai_api_key
export TRANSLATION_MODEL=openai
export GOOGLE_CREDENTIALS_PATH=/path/to/your/credentials.json

3. Run the app

python main.py

Alpaca-Guarani Database Schema Used for this project

Table: alpaca-guarani

This table stores original Spanish texts and their Guarani translations for the Alpaca dataset.

Column Name	Data Type	Constraints	Description
id	int8	PRIMARY KEY	Unique identifier for each entry
original_instruction	text		Original instruction text in Spanish
original_input	text		Original input text in Spanish
original_output	text		Original output text in Spanish
translated_instruction	text		Translated instruction text in Guarani
translated_input	text		Translated input text in Guarani
translated_output	text		Translated output text in Guarani
translated	bool		Flag indicating if the entry has been translated

Notes:

• The id column is used as the primary key and should be automatically incremented.
• The translated column is a boolean flag to easily identify which entries have been translated.
• All text fields (original and translated) are stored as text type to accommodate varying lengths.
• The schema follows the structure of the Alpaca dataset, with separate fields for instruction, input, and output.

Usage:

• When inserting new entries, ensure the id is unique.
• Set translated to false when adding new original texts.
• Update translated to true when the translation is completed.
• Query untranslated entries with: SELECT * FROM alpaca-guarani WHERE translated = false;
• Query translated entries with: SELECT * FROM alpaca-guarani WHERE translated = true;

This schema allows for efficient storage and retrieval of both original and translated texts, facilitating the translation process and progress tracking.