The Largest LLM Benchmarking Suite: MEGAVERSE

Author(s): Dr. Mandar Karhade, MD. PhD.

Originally published on Towards AI.

Now, Benchmarking extends over 81 languages and even 2 multimodal datasets

A quick review of the research published by Microsoft’s Sunayana Sitaram.

As the LLMs become more advanced and more comprehensive, the evaluation frameworks need to keep up with their performance evaluation capabilities over multiple modalities, languages, and variations in the way the evaluation happens.

Microsoft has published the latest Benchmarking Suite: MEGAVERSE

It includes 22 datasets, 81 languages, and 2 multimodal datasets.

Benchmarking tests have been developed for the English language. The largest model that we evaluated, GPT4 (OpenAI, 2023) comes close to but in mostcases do not surpass the performance of SOTA fine-tuned language models such as TULRv6 (Patra et al., 2023). GPT4 performs worse in the non-Latin scripts and on low-resource language.

PaLM2 (Google, 2023)Llama2 (3 variants) (Touvron et al., 2023) andLLaVA-v1.5 (Liuet al., 2023a)GPT4GPT-3.5-TurboLLaVA-v1.5 model (Liu et al., 2023a)two new multilingual multimodal datasets

BIG-bench Srivastava et al. (2023) has 204 tasks, has tests for multiple languages

Holistic Evaluation of Language Models (HELM) Liang et al. (2022), includes (tasks, domains, and languages) and metrics (eg. accuracy, calibration, toxicity), includes 30 language models on 42 scenarios and 7 metrics.

BUFFET (Asai et al., 2023) included 54 languages across 15 datasets

Lai et al. (2023) covering 37 languages across… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI