somosnlp-hackathon-2026/mombeu-LFM2.5-350M
Mombeu-LFM2.5-350M, developed by somosnlp-hackathon-2026, is a 0.35 billion parameter model built upon LiquidAI/LFM2.5-350M. It functions as a narrative insertion layer, designed to append culturally aligned narrative closures based on Guaraní worldview to existing LLM outputs. This model is not a general-purpose chatbot but specializes in enriching text with collective memory and oral tradition, operating locally via `transformers`.
Loading preview...
Overview
Mombeu (meaning "to tell, to narrate" in Guaraní) is a specialized narrative insertion layer built on top of LiquidAI/LFM2.5-350M. Unlike general-purpose LLMs, Mombeu's core function is to take an existing text output from another LLM and, when thematically appropriate, append a narrative closure rooted in the Guaraní worldview. This process aims to provide a channel for Guaraní collective memory and oral tradition to circulate digitally, integrating cultural context into everyday LLM interactions.
Key Capabilities
- Cultural Alignment: Adds narrative continuations based on Guaraní cosmovision to existing text.
- Post-processing Layer: Operates as a secondary step after an initial LLM generates base text, working in the background.
- Local Execution: Designed to run locally via
transformers, compatible with outputs from both local models and external APIs (e.g., OpenAI). - Educational & Cultural Enrichment: Facilitates the transmission of Guaraní heritage and oral tradition in a non-didactic, evocative manner.
Training and Limitations
The model was trained on the somosnlp-hackathon-2026/paraguay-cultural-alignment dataset, undergoing both Supervised Fine-Tuning (SFT) with 2,700 examples and Direct Preference Optimization (DPO) with 1,500 pairs. Its performance is directly influenced by the quality of the base text it receives and the coverage of its Guaraní corpus. Mombeu is not intended for free generation or general instruction-following tasks.