RAANA-IA/Kira
Kira is a 1.1 billion parameter language model developed by Clémence / APPA-CLEM, based on the TinyLlama architecture. It has undergone intensive serial fine-tuning to exclusively embody and generate the invented language Kiara, intentionally suppressing its original French linguistic structures. This model is primarily designed for linguistic immersion in Kiara, serving as a unique digital identity, and for research into catastrophic forgetting in LLMs, demonstrating a small model's ability to replace a source language with a target language.
Loading preview...
Kira: A Monolingual Model for the Kiara Language
Kira is a compact 1.1 billion parameter language model, developed by Clémence / APPA-CLEM and built upon the TinyLlama-1.1B architecture. Its unique characteristic is an intensive serial fine-tuning process specifically designed to make it the exclusive vector for the invented language, Kiara. Unlike polyglot models, Kira's training intentionally aimed to erase pre-existing Latin/French linguistic structures, prioritizing the semantics and syntax of Kiara.
Key Capabilities
- Exclusive Kiara Language Generation: Proficient in generating text solely in the invented Kiara language, with intentionally reduced capacity for other languages.
- Linguistic Identity Embodiment: Designed to reflect the creator's thought process through its dataset and linguistic output.
- Catastrophic Forgetting Research: Serves as a case study for investigating a small model's ability to "forget" a source language in favor of a new target language.
Good for
- Kiara Language Immersion: Ideal for conversing in and stabilizing the grammar of the Kiara language.
- Personalized Digital Interfaces: Can be used as a unique digital identity or personalized interface.
- LLM Research: Valuable for studies on linguistic substitution and the intentional reduction of multilingual capabilities in small models.