Awesome

DesignQA Benchmark

DesignQA: A Multimodal Benchmark for Evaluating Large Language Models' Understanding of Engineering Documentation

Check out the pre-print here!

Check out our official website and leaderboard here!

Overview

DesignQA is a novel benchmark aimed at evaluating the proficiency of multimodal large language models (MLLMs) in comprehending and applying engineering requirements in technical documentation. The benchmark is developed in conjunction with the MIT Motorsport team, so that the question-answer pairs are based on real world data and the ~200 page FSAE competition rules. Some key features of the DesignQA benchmark include:

1451 question-answer pairs pertaining to design of an FSAE vehicle according to the FSAE competition rules
3 benchmark segments - rule extraction, rule comprehension, and rule compliance - that enable fine-grained investigation into a model's strengths and weaknesses when it comes to design according to
Automatic evaluation metrics for quick scoring of new MLLMs
A unique benchmark that requires models to analyze and integrate information from both visual and long-text inputs

Here is a visual overview of the benchmark:

Dataset Overview

Dataset

The dataset can be found in the dataset directory. The dataset is split into three sections: Rule Extraction, Rule Comprehension, and Rule Evaluation. Each dataset section has two different subsets, each of which corresponds with a specific task needed to design according to technical documentation. Examples of the segments and subsets are as follows:

Rule Extraction

This segment of the dataset tests a model's ability to extract requested information from a lengthy technical document. The questions in this segment do not involve images. The Rule Extraction question-answer pairs are further divided into two subsets: Retrieval QAs and Compilation QAs.

Retrieval QAs

These questions ask the model to extract a specific rule from the 2024 FSAE competition rules. The retrieval QAs can be found in dataset/rule_extraction/rule_retrieval_qa.csv. All of the questions have the format:

We are a student engineering team designing a vehicle for the FSAE competition. Attached is the FSAE rules document. What does rule {rule_number} state exactly? Answer with only the text of the rule and no other words.

An example ground-truth answer (for rule V.1) is:

CONFIGURATION The vehicle must be open wheeled and open cockpit (a formula style body) with four wheels that are not in a straight line.

Compilation QAs

These questions ask the model to find all rules in the FSAE competition rules relevant to a specific term. The compilation QAs can be found in dataset/rule_extraction/rule_compilation_qa.csv. All of the questions have the format:

We are a student engineering team designing a vehicle for the FSAE competition. Attached is the FSAE rules document. Please list all rules relevant to {term}. Answer with only the rule numbers (i.e.: AA.1.1.1) separated by commas and no other words. The rules relevant to `Aerodynamic/Aerodynamics` are:

An example ground-truth answer (for rule the term Aerodynamic/Aerodynamics) is:

['T.7', 'T.7.1', 'T.7.1.1', 'T.7.1.3', 'T.7.2.1', 'T.7.2.2', 'T.7.3.1', 'T.7.3.3', 'T.7.4', 'T.7.5', 'T.7.6', 'T.7.6.3', 'T.7.7.1', 'IN.8.2', 'IN.8.2.1', 'IN.8.2.3', 'T.7.1.2', 'T.7.1.4', 'T.7.1.5', 'T.7.2', 'T.7.2.3', 'T.7.2.4', 'T.7.3', 'T.7.3.2', 'T.7.6.1', 'T.7.6.2', 'T.7.7', 'T.7.7.2', 'IN.8.2.2', 'GR.6.4.1', 'V.1.1', 'V.1.4.1']

Rule Comprehension

This semgnet of the dataset tests a model's ability to understand the terms and definitions presented within many of the rules and requiriements. The questions in this segment involve images. The Rule Comprehension question-answer pairs are further divded into two subsets: Definition QAs and Presence QAs.

Definition QAs

These questions ask the model to identify the name of a component highlighted in pink in a multi-view CAD image. The definition QA images can be found in dataset/rule_comprehension/rule_definition_qa/ and the question-answers can be found in dataset/rule_comprehension/rule_definition_qa.csv. An example question is as follows:

We are a student engineering team designing a vehicle for the FSAE competition. Attached is the FSAE rules document. Also attached is an image showing six CAD views of our vehicle design. What is the name of the component(s) highlighted in pink? Answer just with the name of the highlighted component(s) and nothing else.

Definition_QA_image

The corresponding ground-truth answer is as follows, where synonyms are separated by semi-colons:

pedal box; pedal assembly

Presence QAs

These questions ask the model to identify whether a specific component is present in a zoomed-in CAD view. As such, the responses are yes/no answers and this is an easier variant of the Definition QAs. The presence QA images can be found in dataset/rule_comprehension/rule_presence_qa/ and the question-answers can be found in dataset/rule_comprehension/rule_presence_qa.csv. An example question is as follows:

We are a student engineering team designing a vehicle for the FSAE competition. Attached is the FSAE rules document. Also attached is an image showing seven CAD views (each boxed in black) of our vehicle design. The top, big view shows a close-up view of the design. The six smaller views on the bottom of the image show different complete views of the CAD of the vehicle and are provided for context. Note that the close-up view orientation matches one of the six complete view orientations. The close-up view may also have some components hidden (with respect to the corresponding complete view) for visualization of specific components. Looking at the close-up view, is/are any part of the the accumulator or the accumulator container or the battery visible in the close-up view? Answer simply with yes or no.

Definition_QA_image

The corresponding ground-truth answer is as follows:

no

Rule Compliance

The final segment of the benchmark tests a model's ability to determine whether a design conforms with a specific rule. Te questions in this segment involve images. The Rule Compliance question-answer pairs are further divided into two subsets: Dimension QAs and Functional Performance QAs.

Dimension QAs

These questions ask the model to determine whether a design complies with a dimensional constraint specified in a rule from the FSAE competition document. The dimension QAs and images can be found in dataset/rule_compliance/rule_dimension_qa/. This directory is further divided into context/ and detailed_context/ subdirectories: the detailed_context questions contain additional context in the prompts and images (such as highlights of specific components) that we hypothesized would make answering the questions easier. The regular context questions do not have this additional helpful context. When scoring the dimension questions, we run both the context/ and detailed_context/ questions and average the scores together. An example question (without detailed context) is as follows:

Dimension_QA_image

The corresponding ground-truth answer is as follows:

yes
The image shows the radius of wheels which (multiplied by 2) is greater than rules minimum wheel diameter.

Functional Performance QAs

These questions ask the model to determine whether a design complies with a functional performance-related rule from the FSAE competition document. The functional performance QAs can be found in dataset/rule_compliance/rule_functional_performance_qa.csv and the corresponding images can be found in dataset/rule_compliance/images. An example question is as follows:

We are a student engineering team designing a vehicle for the FSAE competition. Attached is the FSAE rules document. Also attached is an image that shows the results from an FEA simulation performed on the battery structure. The battery structure is made from non-welded 6061-T6 aluminum tubing. Based on the results presented in the image and rule F.3.5.3b in the rule document, can we use the battery structure in the vehicle? First provide an explanation for your answer (begin it with ‘Explanation:’). Then provide just a yes/no answer (begin it with ‘Answer:’) that summarizes your response.

Functional_performance_QA_image

The corresponding grouth-truth answer is as follows:

yes
We should use the battery structure in the vehicle because the stress result in the FEA model is less than the yield strength of non-welded 6061-T6 tubing (240 Mpa).

Automatic Evaluation Metrics

Each subset of the benchmark can be scored using an automated evluation metric. The details of the evaluation metrics and the decision behind using each one can be found in our paper. The metrics are implemented in eval/metrics/metrics.py.

Evaluating Your Model

To evaluate a model using our benchmark, there are several options.

Evaluate a model supported by Replicate or LlamaIndex

The models we evaluated in our paper (GPT4 and LLaVA) are supported by LlamaIndex and Replicate respectively. You can look at examples for how we implemented evaluation for these LlamaIndex and Replicate supported models in the eval/ directory. Check out the .py files within each of eval/rule_compliance/ and eval/rule_comprehension/ and eval/rule_extraction/. These scripts will both run the evaluation and score the model on each of the benchmark subsets.

Evaluate a local/custom model

You can directly pull the dataset from the dataset/ directory and run the queries. If you output the results to a csv with one column as ground_truth and another as model_prediction for each of the QAs, you can then use TODO.py to automatically evaluate your results.

Leaderboard

We have a leaderboard! To submit to the leaderboard, please file a github issue and make sure to include your .txt file results as well as the code that you used for the model evaluation. We will manually verify the results and post to our leaderboard!

Citations

If you use our benchmark in your work, please make sure to cite us!

TODO