The project is an Apache 2.0 licensed open source project available on Github: https://github.com/open-llm-initiative/llm-challenge. And the challenge is hosted on AWS as a single-page web app, where users see greeting text, followed by a randomly selected prompt and a LLM response, which they must rate on a likert scale of 1-5 (or yes/no rating) that matches the task represented in the prompt.

The study uses pre-generated prompts across popular real-world uses cases like information extraction and summarization, creative tasks like writing a blog post or story, problem solving task like getting central ideas from a passage or writing business emails or brainstorming ideas to solve a problem at work/school. And to generate responses of varying quality the study uses the following OSS LLMs: Qwen 2-0.5B-Instruct, Qwen2-1.5B-Instruct, gemma-2-2B-it, Qwen2-7B-Instruct, Phi-3-small-128k-instruct, Qwen2-72B and Meta-Llama-3.1-70B. And for proprietary LLMs, we limited our choices to Claude 3 Haiku, Claude 3.5 Sonnet, OpenAI GPT 3.5-Turbo and OpenAI GPT4-o.

Today, LLM vendors are in a race with each other to one-up benchmarks like MMLU, MTBench, HellowSwag etc - designed and rated primarily by human experts. But as LLMs get deployed in the real-world for end users and productivity workers, there hasn't been a study (as far as we know) that helps researches and developers understand the impact of model selection as perceived by end users. This study aims to get valuable insights to incorporate human-centric benchmarks in building generative AI applications and LLMs

If you want to contribute to the AI community in an open source way, we'd love if you can take the challenge. We'll publish study results in 30 days on Github.