PCA-Bench/PCA-Bench-V1

PCA-Bench

PCA-Bench is an innovative benchmark for evaluating and locating errors in Multimodal LLMs when conducting embodied decision making tasks, specifically focusing on perception, cognition, and action.

Release

• [2024.02.15] PCA-Bench-V1 is released. We release the open and closed track data in huggingface. We also set an online leaderboard accepting users' submission.
• [2023.12.15] PCA-EVAL is accepted to Foundation Model for Decision Making Workshop @NeurIPS 2023. PCA-Evaluation tool is released in github.

Leaderboard

Leaderboard with Full Metrics

Submit Results

📢 For close track evaluaiton and PCA-Evaluation, please follow this file to organize your model output. Submit Six JSON files from different domains and different tracks, along with your model name and organization to us via email. Ensure you use the dataset's provided prompt as the default input for fair comparison.

We will send the PCA-Eval results of your model to you and update the leaderboard.

We provide sample code to get the six json files. User only needs to add your model inference code:

# Sample code for PCA-Eval
from datasets import load_dataset
from tqdm import tqdm
import json
import os

def YOUR_INFERENCE_CODE(prompt,image):
    """Simple single round multimodal conversation call.
    """
    response = YOUR_MODEL.inference(prompt,image)
    return response

output_path = "./Results-DIR-PATH/"
os.mkdir(output_path)

dataset_ad = load_dataset("PCA-Bench/PCA-Bench-V1","Autonomous Driving")
dataset_dr = load_dataset("PCA-Bench/PCA-Bench-V1","Domestic Robot")
dataset_og = load_dataset("PCA-Bench/PCA-Bench-V1","Open-World Game")

test_dataset_dict = {"Autonomous-Driving":dataset_ad,"Domestic-Robot":dataset_dr,"Open-World-Game":dataset_og}
test_split = ["test_closed","test_open"]
test_domain = list(test_dataset_dict.keys())

for domain in test_domain:
  for split in test_split:
    print("testing on %s:%s"%(domain,split))

    prediction_results = []
    output_filename = output_path+"%s-%s.json"%(domain,split)
    prompts = test_dataset_dict[domain][split]['question_prompt']
    images = test_dataset_dict[domain][split]['image']

    for prompt_id in tqdm(range(len(prompts))):
        user_inputs = prompts[prompt_id] # do not change the prompts for fair comparison
        index = prompt_id
        image = images[prompt_id]

        outputs = YOUR_INFERENCE_CODE(user_inputs,image)

        prediction_results.append({
            'prompt': user_inputs,
            'model_output': outputs,
            'index': index,
        })

    with open(output_filename, 'w') as f:
        json.dump(prediction_results, f, indent=4)

# submit the 6 json files in the output_path to our email

You could also simply compute the multiple-choice accuracy locally as a comparison metric in your own experiments. However, in the online leaderboard, we only consider the average action score and Genuine PCA score when ranking models.

For more information, refer to the offical github repo