Frostie08/Luma-base
Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Feb 28, 2026Architecture:Transformer Warm

Luma-base by Frostie08 is a 4-billion parameter causal language model, based on the Qwen3-4B architecture, specifically pre-trained for Haitian Creole (Kreyòl Ayisyen). It features a 32768-token context length and is optimized for tasks like speech-to-text correction, translation, and text generation in Haitian Creole. The model achieved a final validation loss of 1.9252 and a perplexity of approximately 6.8, demonstrating high confidence in Haitian Creole word prediction.

Loading preview...

Luma-base: Haitian Creole Foundation Model

Luma-base, developed by Frostie08, is a 4-billion parameter causal language model built upon the Qwen3-4B architecture. Its primary distinction lies in its specialized and extensive pre-training on a high-quality Haitian Creole corpus, kani-pretrain, making it a dedicated resource for the language.

Key Capabilities & Features

  • Haitian Creole Specialization: Deeply understands the nuances, grammar, and cultural context of Haitian Creole (ht-HT).
  • Efficient Training: Utilized the Unsloth library for training, ensuring efficiency and mathematical precision.
  • Performance Metrics: Achieved a final validation loss of 1.9252 and a perplexity of ~6.8, indicating strong language modeling capabilities for Haitian Creole.
  • Robust Architecture: Based on the Qwen3-4B model, providing a solid foundation for language tasks.

Ideal Use Cases

Luma-base is designed to serve as a core engine for various applications requiring high-quality Haitian Creole processing:

  • Speech-to-Text (STT) Correction: Enhancing accuracy in transcribing spoken Haitian Creole.
  • Machine Translation: Improving translation quality between Haitian Creole and other languages.
  • Text Generation: Creating coherent and contextually relevant text in Haitian Creole.
  • Language Research: Providing a robust base model for further research and development in Haitian Creole NLP.