In our experiments, we assess multiple language models based on their architectural designs and training datasets, focusing on their multilingual capabilities and performance.
The benchmarking of models such as GPTNEO and BLOOM reflects divergent training strategies; primarily monolingual versus multilingual datasets significantly influence their capabilities in translation tasks.
We leverage FLORES datasets to systematically evaluate bilingual generation tasks, utilizing BLEU scores as the metric for assessing translation accuracy across the tested models.
Prompt design plays a crucial role in model outputs; by using neutral delimiters, we mitigate biases associated with input instructions, leading to more consistent evaluations.
[
Collection
]
[
|
...
]