Llama Impact Hackathon

Start

LRL USFCA

karimnihal

victorweiData

This project, Llama-Chain-Translate, adapts the GEMBA framework for improved machine translation evaluation and enhancement, specifically targeting low-resource languages (LRLs). The application leverages chain-of-thought prompting techniques using open-source models hosted via Together.ai, enabling reproducibility and transparency. First, it assigns a large language model (LLM) as an evaluator, tasked with assessing the fluency and adequacy of translations. Then, a separate LLM is designated as a translator, using evaluation feedback to produce an improved translation. Future directions include fine-tuning the model, domain adaptation, back-translation, noisy channel reranking, and specific tokenization techniques for LRLs to tackle data scarcity and other LRL challenges.

Builds on the GEMBA repository to improve machine translations using chain of thought techniques. It assign one LLM as an evaluator and another as a translator, refining translations based on initial evaluations for enhanced accuracy.

"LRL USFCA team on Llama Impact Hackathon Hackathon"

Team Idea

Submission