Vibe-Eval

A benchmark for evaluating multimodal chat models, including especially challenging examples.

[Link to paper] [Blogpost] [🤗 Dataset]

Dataset

The dataset including all images can be downloaded in the Releases page of this repo.

The dataset is stored as a JSONL file: data/vibe-eval.v1.jsonl. Each example has the following fields:

• example_id: a unique ID for the example
• category: the category that this example belongs to, either difficulty-normal or difficulty-hard
• prompt: the user prompt
• reference: a golden reference answer for the prompt
• media_filename: the name of the file in the dataset
• media_url: a URL where the file is hosted publicly

Running the evaluation

To run the evaluation, use evaluate.py as follows:

python evaluate.py generations.jsonl -o out.jsonl

(you will have to install a couple of requirements, including the Reka API package with pip install -r requirements.txt)

The generations.jsonl is expected to contain model generations. It should be a JSONL file where each line is a JSON object with keys "generation" and "example_id" (matching the dataset).

This will output detailed results to out.jsonl and will also print a table of final results to stdout.

Leaderboard 🏆

Vibe-Eval Score (%)

Model	all	hard	normal
GPT-4o	63.1	54.1	68.1
Gemini Pro 1.5	60.4	53.0	64.8
GPT-4V	57.9	46.0	64.9
Reka Core	53.7	38.2†	62.8
Claude Opus	52.8	41.8	59.2
Reka Flash	52.2	39.2	59.9
Claude Sonnet	52.1	39.7	59.5
Claude Haiku	49.8	38.5	56.4
Llava-1.6-34b	48.6	39.9	53.7
Reka Edge	45.4	32.2	53.1
Llava-1.6-7b	43.7	35.3	48.6
Idefics-2-8b	40.0	32.2	44.6
Idefics-1-80b	36.0	32.1	38.3
Fuyu-8b	30.8	23.4	35.2

† Note we expect the results of Reka Core to be worse on the hard-set, as these are, by their very definition, prompts that Core cannot solve.

Citation

@article{padlewski2024vibeeval,
  title={Vibe-Eval: A hard evaluation suite for measuring progress of multimodal language models},
  author={Piotr Padlewski and Max Bain and Matthew Henderson and Zhongkai Zhu and Nishant Relan and Hai Pham and Donovan Ong and Kaloyan Aleksiev and Aitor Ormazabal and Samuel Phua and Ethan Yeo and Eugenie Lamprecht and Qi Liu and Yuqi Wang and Eric Chen and Deyu Fu and Lei Li and Che Zheng and Cyprien de Masson d'Autume and Dani Yogatama and Mikel Artetxe and Yi Tay},
  journal={arXiv preprint arXiv:2405.02287},
  year={2024}
}